[预印本] ISO 11783网络上机器可读的操作手册与离线多语种大模型辅助（Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance）

Machine-Readable Operator Manuals over the ISO 11783 Network for
Offline, Multilingual LLM Assistance
Timo Oksanena,b,∗, Julius Guna
aChair of Agrimechatronics, Technical University of Munich, Germany
bMunich Institute of Robotics and Machine Intelligence (MIRMI), Germany
ARTICLE INFO
Keywords:
large language models
ISOBUS
natural language processing
retrieval-augmented generation
user manuals
Abstract
If agricultural machinery is to support intelligent, offline assistance, then the system must access
technical documentation dynamically. In this paper, the Semantic Synchronization of operator
manuals over ISO 11783-based CAN buses is analyzed, defined as the process that brings the local
knowledge base of a tractor-side assistant into agreement with the operator manual of the connected
implement.DatatransferperformancewasbenchmarkedusingMarkdown,JSON,andXML;various
open-weight Large Language Models (LLMs) were also evaluated in cross-lingual query tasks. The
automatic evaluation of the answers was validated against four human raters. A complete manual
was transferred over a physical ISO 11783 network in 31 to 89 s, depending on format and transport
settings, which is comparable to the object pool upload that operators already wait for at the first
coupling of an implement. Markdown produced the smallest files, while the question-answering
performance was comparable across the three formats for most models. With retrieval-augmented
generation, locally run models answered most questions on the English manual correctly, with no
advantageofthelargestmodelover7Band8Bmodels,butperformedworseonthefiveotherlanguage
editions. Keyword retrieval failed across languages, whereas semantic retrieval provided the cross-
lingual capability. The exchange requires only a new parameter group that uses the existing request
and transport mechanisms of ISO 11783, which makes it a candidate for standardization. The study
is a proof of concept with inference on PC-grade hardware; deployment on embedded agricultural
hardware remains future work.

1. Introduction
   Integrating Large Language Models (LLMs) into agri-
   culturalmachineryrequiresarobustsolutionforofflinedata
   access.Printedoperatormanualscannotprovidein-situsup-
   portinthetractorcabinbecausetheyaredifficulttonavigate
   during operations. Although researchers have investigated
   theseLLMsforofficeandmobilesupport,applicationsnow
   extend to embedded systems. To integrate these systems
   into agriculture, however, we must navigate the specific
   constraints of the standardized vehicle network.
   Agricultural machinery in Europe features multi-brand
   systems where farmers connect implements to the tractor
   in a plug-and-play manner using ISO 11783 (ISO, 2017).
   Farmerssupplementtheconnectedsystemwithretrofitelec-
   tronics, such as GNSS receivers and user terminals. ISO
   11783 (ISOBUS) is the standard that interconnects the trac-
   tor, implement, user interface, farm management system,
   and precision farming task controller. This interconnected
   environment provides the necessary infrastructure for ad-
   vanced data exchange.
   In this paper, we explain how a local LLM-based as-
   sistant in an ISO 11783 system accesses textual data from
   other subsystems in an open network. Specifically, we ex-
   amine machine-readable digital operator manuals stored in
   animplement’sElectronicControlUnit(ECU)andaccessed
   by a tractor terminal in the tractor cabin. Access to these
   data is crucial in rural areas where internet connections are
   unreliable. Storing the manual in the implement ECU keeps
   ∗Corresponding author: timo.oksanen@tum.de
   ORCID(s):
   it with the machine throughout its lifecycle. This study is
   a proof of concept: we evaluate the transfer on a physical
   ISO11783networkandthequestionansweringwithlocally
   run open-weight LLMs on PC-grade hardware, as a step
   toward deployment on resource-constrained ECUs. Unlike
   autonomous driving, where AI actuates the vehicle, the
   proposedsystemoperatesstrictlyasaninformationretrieval
   layer (Level 0 autonomy), which places the human operator
   as the sole decision maker.
   The ISO 11783 series standardizes the user interface
   (Virtual Terminal, ISO 2018b), task data exchange (Task
   Controller, ISO 2015) and diagnostics limited to version
   information, capabilities and I/O errors (ISO, 2019a), but it
   provides no mechanism by which an implement can supply
   its technical documentation to a local assistant on the trac-
   tor side, and existing retrieval-augmented systems rely on
   centrallymaintaineddocumentcollections(Ateşetal.,2025;
   Koopman et al., 2022). A standardized exchange is needed,
   which would allow tractor manufacturers to build their own
   assistantsandimplementmanufacturerstoprovidetheirown
   serialization of the manual. Following the layered ISO/OSI
   architecture of ISO 11783, such an exchange should be
   added at the application layer as a new parameter group
   that uses the existing network management, request and
   extended transport mechanisms unchanged, rather than as a
   proprietary extension. We defineSemantic Synchronization
   as the process that brings the local knowledge base of the
   tractor-side assistant into agreement with the manual of the
   connected implement, as in data synchronization: its input
   is the versioned manual stored in the implement ECU, its
   T. Oksanen and J. Gun:PreprintPage 1 of 18
   arXiv:2610.31863v1 [cs.CL] 22 Oct 2026

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
output a complete local copy together with the retrieval
index built from it; it succeeds when the manual is received
completely, its version matches the one announced by the
implement and the index is built from it. In this study,
the transfer and the indexing were implemented and evalu-
ated, whereas the version check, the first-coupling protocol,
the background image transfer and the source-verification
interface are proposed design elements. Building on our
preliminary study (Gun and Oksanen, 2025), we reuse its
manual,question-answerdataset,promptsandretrievalcon-
figuration; the transfer over ISO 11783, the format com-
parison, the six language editions and the current models
are new. This ManualSync project aims for standardized
exchange of machine-readable operator manuals over ISO
11783 for offline, multilingual LLM assistance. We address
threeresearch questions:

1. Can a complete operator manual be transferred over
   ISO 11783 with the existing request and extended
   transport mechanisms within a practical time at the
   first coupling of an implement?
2. HowdoMarkdown,JSONandXMLcompareinsize,
   transfertime,tokencountandquestion-answeringper-
   formance?
3. Can locally run open-weight LLMs answer operator
   questionsfromthetransferredmanual,andalsoacross
   languages? What does each retrieval component con-
   tribute?
4. Motivation
   We propose a self-describing architecture to bridge the
   gapstatedabove.Inourdesign,theimplementstoresitsown
   documentation in a structured format and transmits these
   datatothelocalLLMofthetractorviatheISOBUSnetwork
   (Figure 1). This data flow allows the operator to maintain
   conversational access to safety information regardless of
   cellular coverage.
5. State of the Art
   3.1. Applications of LLMs
   Researchers have recently integrated LLMs into
   automotive and heavy-equipment domains to enhance
   user interaction. In autonomous driving, LLMs improve
   decision-making by combining multimodal sensor inputs
   (Lietal.,2024),whileintheintelligentcockpit,theyenable
   natural-language control of vehicle functions (Jafarnejad
   et al., 2024). These models predict driver intent accurately
   and provide a superior user experience compared to
   conventional menu-driven interfaces (Chen et al., 2025). In
   the specific context of agriculture, Johnson and Wilson
   (2024) showed that multi-round prompt engineering
   significantlyimprovesmachinerymanagement.Multimodal
   assistantssuchasAgri-LLaVAaddressagronomicquestions
   on pests and diseases (Wang et al., 2024).
   Recentstudies alsoinvestigate architectureswherelarge
   language models interact with structured technical infor-
   mation or domain knowledge sources in industrial systems
   (Ateş et al., 2025; Rohmadin et al., 2025). Although these
   works explore LLM-driven knowledge interfaces, our work
   distinguishes itself by enabling such functionality directly
   over the ISO 11783 (ISOBUS) machinery network. The do-
   main has specific requirements for hot plug-and-play func-
   tionality, making ISOBUS the fundamental constraint for
   our system design.
   3.2. ISO 11783 Transfer Mechanism
   To move the data required by these models, it is
   necessary to look to the existing transport layers of the
   vehicle network. The ISO 11783 standard series provides
   the Extended Transport Protocol (ETP) for transmitting
   extra-large data messages over the CAN bus. The classic
   Transport Protocol (TP), originally developed jointly for
   the SAE J1939 and ISO 11783 series, limits message
   transmission to 1785 bytes. ETP overcomes this constraint
   by supporting messages up to approximately 117 million
   bytes through enhanced packet addressing capabilities.
   ETP uses two distinct Parameter Groups: ETP.CM
   (Connection Management, PGN=51200) for control
   messages and ETP.DT (Data Transfer, PGN=50944) for
   payload transmission (ISO, 2018a).
   The protocol manages the connection through a spe-
   cifichandshakesequence.ETPoperatesthroughconnection-
   orientedsessionswheretheoriginatorinitiatesthesessionby
   sending a Request to Send (RTS) message that specifies the
   total message size and the target PGN. If ready, the receiver
   responds with a Clear to Send (CTS) message to indicate
   reception capacity. Although 16 is the industry default, this
   capacity supports 255 packets per burst. Each data packet
   contains seven bytes of payload and the sequence number.
   If the receiver detects packet loss, it requests retransmission
   using Data Packet Offset commands. An End-of-Message
   Acknowledgement confirms session completion, whereas
   abort messages terminate a session abnormally.
   This protocol manages the 8-byte limitation of the CAN
   bus, making it the primary candidate for transferring large
   documentation files.
   3.3. Cross-lingual Capability in Machinery
   Because agricultural machinery operates across diverse
   linguistic regions, manufacturers must provide operator
   manualsinmultiplelanguages.Chuaetal.(2025)evaluated
   LLMs on cross-lingual tasks and found that although
   modelsshowpromisingsurface-levelabilities,theystruggle
   with deep cross-lingual knowledge transfer. Their work
   demonstrated that fine-tuning on mixed-language data
   effectively reduces these performance gaps.
   Recent agricultural studies confirm the value of mul-
   tilingual access. Kaur et al. (2025) generated multilingual
   synthetic datasets from agriculture-specific documents and
   achieved improvements in factuality by fine-tuning LLMs
   for question answering across English, Hindi, and Punjabi.
   T. Oksanen and J. Gun:Preprint Page 2 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
ECU
Implement
(data source)
ECU
Tractor
(compute node)
ISO 11783 network
Operator manual assistant Manual of the connected implement 14:32
Operator manual (English) › Page 43 › Vanes Source 1 / 3
The spreader is provided with 12 vanes of equal length and 4 shorter.
Grease the threaded rods of the vanes.
Hook the short vanes into the correct position (see illustration).
Then place the other two vanes into the other positions.
Press the vanes in well and put the lock nut in position.
Place the vanes on the other side in the same manner.
Fasten the lock nuts (50 Nm)
Even wear of the vanes
The long vanes are positioned in groups of three.
The first vane considered in the direction of rotation of the disc will wear out
most quickly.
TIP Exchange the first vane after the fixed vane, and the short vane, over a
period of time.
After a subsequent period, exchange the foremost vane at that time with
the central vane.
Today, 14:31
What torque (in Nm) should be
applied to the vane lock nuts?
Operator
The vane lock nuts should be fastened to
50 Nm.
Verify: please check the highlighted passage in
the manual.
Verify source
Ask the manual…
Tractor terminal
Operator

1. Uploads
   manual
2. Local LLM
   inference
3. Natural
   language query
4. Verifiable
   answer
   Figure 1:Conceptual architecture for offline Semantic Synchronization.
   Similarly, Samuel et al. (2025) demonstrated that combin-
   ing LLMs with RAG frameworks achieves high accuracy
   for agricultural knowledge sharing. These findings suggest
   that an LLM with cross-lingual capabilities could answer
   user queries in one language using documentation stored in
   another, reducing the need to store multiple translations in
   the resource-constrained memory of an ECU.
   3.4. Retrieval-Augmented Generation (RAG)
   Retrieval-Augmented Generation (RAG) offers a solu-
   tion to the limited knowledge base of pretrained LLMs.
   Vizniuk et al. (2025) identify RAG as a key approach for
   overcomingthelimitedknowledgebaseofpretrainedLLMs
   in agricultural decision-support systems. At the retrieval
   level, Saha et al. (2024) match user questions against ques-
   tionsgeneratedfromthetargetcorpustoreduceinformation
   dilution and hallucination in RAG-based question answer-
   ing.
   Several systems already follow this pattern. AgAsk
   (Koopman et al., 2022) answers farmers’ natural-language
   questions by retrieving relevant passages from scientific
   documents, effectively implementing an early RAG-style
   pipeline for agriculture-focused QA. PooRaa-Agri KG
   (Sivakumar et al., 2023) uses a knowledge graph and a
   multilingual query interface to return concise answers in
   English and Hindi, separating retrieval over a structured
   domain representation from the answer layer. AgroLLM
   (Samuel et al., 2025) explicitly combines a vector database
   (FAISS) with LLMs in a RAG framework. The authors
   report substantial accuracy gains over non-RAG baselines
   across several agricultural subdomains.
   At the architectural level, Akbar et al. (2025) review
   RAG-driven memory designs for conversational LLMs.
   They highlight how vector databases can integrate long-
   term domain knowledge with user-specific data. In parallel,
   recent work shows that modern tractors and implements
   already expose rich digital traces via CAN bus and GNSS
   (da Silva et al., 2019; Bettucci et al., 2024). These strands
   suggest that RAG architectures are well suited for linking
   machinery data to technical documentation.
   3.5. Large Context Window
   Recent advancements in LLMs allow models to pro-
   cess entire documents simultaneously, but this capability
   has limits. Models like GPT-OSS 20B and Gemma 3 12B
   now support context windows of up to 128k tokens. Liu
   etal.(2023),however,describethelost-in-the-middleeffect,
   where models fail to retrieve information located in the
   middle of long contexts. It was found that small models
   performbetterifthesystemusesRAGtosupplyonlyasmall,
   relevant context (Gun and Oksanen, 2025). This approach
   avoids the degradation associated with processing massive
   documents directly.
5. Materials
   4.1. Operator Manual Structure
   Under EU Machinery Directive 2006/42/EC (EU,
   2006), agricultural machinery manufacturers must provide
   a comprehensive operator manual to meet health and
   safety requirements. Manufacturers typically achieve
   compliance through the harmonized standard EN ISO
   4254-1 (ISO, 2013), which sets requirements for the
   content of the operator manual. These requirements include
   mountinganddismountingprocedures,tractorcompatibility
   specifications, stability information, control descriptions
   with standardized symbols, ergonomic adjustments,
   T. Oksanen and J. Gun:Preprint Page 3 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
starting and stopping procedures, and emergency exits.
Additionally, the manual must cover moving-parts
precautions, maintenance procedures, noise and vibration
emission values, energy hazards (battery and fuel), lifting
points, and power take-off shafts.
These regulations create structural homogeneity across
operator manuals from different manufacturers. Because
regulationsmandateseveralchaptersintheoperatormanual,
thedocumentorganizationispredictable,evenifnotunified.
This standardization driven by legal compliance creates an
interesting foundation for machine-readable documentation
systems because information categories may allow auto-
mated content extraction, classification, and analysis. We
selected the operator manual of an agricultural fertilizer
spreader because its structure is shared by manuals of sim-
ilar machines under the scope of this regulation. Fertilizer
spreadershavetheirownstandardforsafetyandhealth,ISO
4254-8 (ISO, 2018c).
4.2. Selected Operator Manual
Our test case is the operator manual for the Kverneland
Exacta-TLX Geospread GS3, a mechatronic fertilizer
spreader.Wechosethismanualbecauseourfamiliaritywith
the machine aided in creating the handcrafted QA set. This
machine also met our requirement for operator manuals
publicly available without authentication in numerous
European languages.
Themanualswereusedintheirofficialeditions:English
(Kverneland Group, 2021e), French (Kverneland Group,
2021d), German (Kverneland Group, 2024), Italian (Kver-
neland Group, 2021c), Dutch (Kverneland Group, 2021a),
and Spanish (Kverneland Group, 2021b). We verified that
all six editions refer to the same machine and software
version.Each165-pagemanualpossessesanidenticallayout
across languages; the Markdown version contains approx-
imately 61k tokens in English and 83k to 89k tokens in
theotherlanguages(Qwen3-Embedding-8Btokenizer).The
numberofwordsvariesbylanguage:English(41094),Dutch
(37972), French (46047), German (38041), Italian (43201),
and Spanish (45751).
Howimagesaffectdataexchangewasalsoexplored.Be-
cause the manual was not designed for machine readability,
one representative image was evaluated for benchmarking.
Thisimage,showninFigure2,isfrompage86ofthemanual
and illustrates a graphical table of PTO speeds for various
boundary spreading settings. Most graphics in this manual
follow a similar style.
Because ISO 11783-6 already provides a presentation
layer for picture graphics, the same compression method
was used. The sample image is natively 578 pixels wide
and 610 pixels high, resulting in 352580 bytes in the 256-
indexed palette used in ISO 11783-6. Two compression
methods can be enabled for this image: reduction to a 16-
indexed color palette and run-length encoding (ISO 11783-
6,sectionB.12).Run-lengthencodingcompressesrepeating
sequences of the same value with two-byte pairs. We se-
lected this method because the ISO 11783 Virtual Terminal
Figure 2:Representative image from the operator manual
(page 86) illustrating PTO speeds for boundary spreading.
server software implements it, and it is industry-proven and
lightweight. Together, the 16-indexed color palette and run-
length encoding provide a compression baseline for the
following benchmark.
4.3. QA Test Dataset
TheQAdatasetfromourpreviousstudy(GunandOksa-
nen,2025),whichcomprises108questionscuratedfromdo-
main expertise, was used. The complete dataset is available
in the associated code repository. The dataset is balanced
with 54 answerable questions targeting specific operational
and safety information (e.g., torque specifications, mainte-
nance intervals) and 54 unanswerable questions designed to
test model hallucination tendencies.
Unanswerable questions are semantically plausible but
cannot be answered from the manual content (e.g., "What
is the minimum baud rate for RS 232 connection?"). This
design enables us to evaluate both retrieval accuracy and
the model’s ability to appropriately decline answering
when information is absent—a critical capability for safety-
relevant applications. Following established benchmarks
like XTREME (Hu et al., 2020), all questions were posed
in English regardless of the document language to isolate
cross-lingual retrieval capability. Example questions from
both categories are provided in Appendix A.
5. Methods
We converted the publicly available PDF manuals into
Markdown, JSON, and XML using the Docling library
(Auer et al., 2024) and developed a wrapper to enable page-
wiseconversion. Thewrapper convertseach PDFonce with
thedefaultDoclingPDFpipeline(Doclingversion2.26)and
serializes the recognized items page by page into the three
formats:sectionheaders,textandlistitemsastext;tablesas
Markdowntables,JSONobjectsorfixed-widthtextinXML;
and pictures, including safety pictograms, as placeholders
with their caption where one was recognized. Page headers
andfooterswerenotincluded.Theconvertedfileswereused
without manual editing and were stored as UTF-8 text with
Windows line endings (CRLF), as transferred.
T. Oksanen and J. Gun:Preprint Page 4 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
5.1. Semantic Synchronization
To evaluate the transfer of manual data over the vehicle
network, we established a test setup using the ISO 11783
Extended Transport Protocol (ETP). We used a temporary
proprietary PG number (65280) to prove the concept, to-
gether with a standard request message (PGN 59904) and
the existing ETP mechanism. The priority of messages was
set to 7 (lowest), which is the default for ETP.
We created the test setup with a physical CAN bus
and two connected computers. One software application
acted as the operator manual server and the other as the
operator manual client. During the test, the physical CAN
bus backbone also ran other control functions, namely an
AEF-certifiedcommercialVirtualTerminal,taskcontroller,
and file server, to provide a realistic plug-and-play system.
Without the manual transfer, the average background bus
load from these control functions was below 2%. At 250
kbit/s (ISO, 2019b), the bus carries approximately 1900
frames per second (131 bits per 8-byte extended frame,
excluding bit stuffing). This background corresponds to the
condition at the first connection of an implement, when the
systemisstationaryandtheoperatorwaitsfortheobjectpool
upload (ISO, 2018b). We aimed to ensure that the normal
operationoftheplug-and-playnetworkstaysfullyfunctional
despite the additional burst of this data exchange.
Intheproposeddesign,thesystemdoesnotperformthis
transfer during every Key-On cycle. Instead, it uses a first-
couplingprotocol:themanualistransferredonlywhenanew
implement is attached for the very first time and is stored
in the tractor’s non-volatile memory. To handle implement
firmware updates that might alter the manual, the system
employsaversioncheck.TheECUcheckstheversionofthe
manual on every startup; a transfer only occurs if it differs
from the cached version. Subsequent connections require
only this check. This follows the existing pattern of ISO
11783-6, in which the Virtual Terminal can store the object
poolofanimplementunderaversionlabel,sothatthepoolis
uploaded only at the first connection or after a change (ISO,
2018b); device descriptions of the task controller, which
are transferred at the same time, take precedence over the
manual.
For Semantic Synchronization, the system serialized
each operator manual into a byte array. This serialization
preservesencodingbutraisesthequestionoftheappropriate
structural format: JSON, XML, or Markdown. Because we
havesixlanguagesandthreestructuralformats,wetested18
combinations.Wemeasuredthetimerequiredtotransferthe
complete manual in different formats to determine the most
bandwidth-efficient solution. The system also measured the
busloadfromtheserversidebyaccessingthehardware-level
diagnostics counters, which account for bit stuffing.
5.2. RAG Configuration
We used a Hybrid RAG retrieval strategy. We detail the
ablation study of the retrieval configuration in Appendix E.
ThisstrategycombinesBM25keywordsearchwithsemantic
retrieval using theQwen3-Embedding-8Bmodel. To merge the
results, we used Reciprocal Rank Fusion (RRF) (Cormack
et al., 2009), which produces a robust final ranking by
prioritizing chunks that rank highly in both methods. We
split documents into 200-token segments with a 100-token
overlapandretrievedthetopthreechunksperquery.Chunks
were created with the tokenizer of the embedding model,
whichranontheCPUin32-bitfloatingpoint.BM25(Okapi
variant) used lowercase whitespace tokens; the semantic
search used ChromaDB with L2 distance on normalized
embeddings, which ranks identically to cosine similarity.
Each retriever returned ten candidates, which were fused
with RRF (𝑘= 60); the top three fused chunks were passed
to the model. This configuration was chosen on the basis of
our preliminary study (Gun and Oksanen, 2025) and is re-
evaluated in Appendix E.
5.3. Models
We evaluated a range of open-weight LLMs deployed
locallyviaOllama(Table1).Allmodelswereevaluatedwith
atemperatureof0,withoutafixedseed,withanoutputlimit
of1024tokensandwiththedefaultreasoningsettingofeach
model. Confidence intervals were estimated by question-
levelbootstrapresampling(2000resamples,stratifiedbyan-
swerable and unanswerable questions; questions resampled
as clusters when editions were pooled), and paired compar-
isons used a paired bootstrap with Holm correction within
each family of comparisons (formats; context conditions)
(Appendix C).
Modelsizesrangefrom1.5Bto120Bparameters,which
enablesanalysisofthetrade-offbetweenmodelcapacityand
performance.Allmodelswereexecutedwith4-bitquantiza-
tion (Q4_K_M; MXFP4 for GPT-OSS), which reduces the
model weights of the 8B-parameter models to about 5 GB.
For automatic evaluation, we used an LLM-as-a-judge
framework (Zheng et al., 2023) with GPT-OSS 20B as the
evaluator across all experiments, ensuring consistency. The
judgecomparedmodeloutputsagainstground-truthanswers
using the prompts in Appendix B. If the judge returned no
valid verdict (147 of 15552 items, 0.9%; mostly because its
reasoning reached the output limit), the call was repeated
with identical settings (123 items resolved) and, for the
remaining 24 items, with the reasoning effort set to low. A
verdict of yes counted as a true positive for answerable and
as a true negative for unanswerable questions, and a verdict
of no as a false negative or a false positive, respectively;
answers written in the language of the manual edition in-
stead of English counted as incorrect. The language of each
answerwasdeterminedwithanautomaticlanguagedetector
(lingua); answers mixing English with the language of the
manual (226) were scored by the verdict of the judge. The
verdictsofthejudgewerevalidatedagainstfourindependent
human raters, with whose majority the judge agreed for
93.5% of the items (Cohen’s𝜅= 0.87); its errors were
mainly rejections of acceptable answers, so the reported F1
scores are conservative (Appendix F).
T. Oksanen and J. Gun:Preprint Page 5 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
Table 1
Overview of LLMs evaluated in this study.
Model Size Ollama tag Parameters Quantization
GPT-OSS 120B gpt-oss:120b 116.8B MXFP4
GPT-OSS (judge) 20B gpt-oss:20b 20.9B MXFP4
Gemma 3 12B gemma3:12b 12.2B Q4_K_M
Qwen 3 8B qwen3:8b 8.2B Q4_K_M
DeepSeek-R1 8B deepseek-r1:8b 8.2B Q4_K_M
Qwen 2.5 7B qwen2.5:latest 7.6B Q4_K_M
Llama 3.1 8B llama3.1:8b 8.0B Q4_K_M
Llama 3.2 3B llama3.2:3b 3.2B Q4_K_M
DeepSeek-R1 1.5B deepseek-r1:1.5b 1.8B Q4_K_M
5.4. Formatting Analysis
To evaluate the efficiency of data transfer and inges-
tion, we selected three formats representing distinct data
interchangeparadigms.Weexplicitlyexcludedunstructured
plain text because it fails to preserve the document hierar-
chy (headers, lists, warnings) that accurate RAG retrieval
requires. We also excluded HTML because of the security
risksassociatedwithrenderingwebcontentinsafety-critical
vehicle displays. Furthermore, we excluded compressed of-
fice formats (e.g., DOCX) because the decompression pro-
cess consumes excessive CPU cycles during the critical
machine startup phase.
Weexcludedbinaryserializationformats(e.g.,Protobuf)
althoughtheysavebandwidth.Wearguethatbinaryformats
introduce a schema coupling dependency: they require both
thetransmitter(implement)andreceiver(tractor)topossess
identical schema definitions (e.g., .proto files) to decode the
payload. This dependency violates the plug-and-play core
principle of ISOBUS; a tractor would require a firmware
update to read documentation from a new implement. In
contrast, the text-based formats below are self-describing
and require no prior schema exchange.
•XML (The Legacy Standard): XML represents the
incumbent standard for agricultural and automotive
data(e.g.,ISO11783-10TaskData,ODX).TheXML
files use a flat structure of page, text, table and image
elements without item types.
•JSON (The Web/API Standard): JSON represents
themodernstandardfortelematicsandcloudconnec-
tivity. The JSON files list the items of each page as
objectswiththeiritemtype(e.g.,sectionheader,text,
list item) and are indented by two spaces.
•Markdown (The LLM-Native Standard): We se-
lected Markdown as the representative stream-based
format. It offers a content-first approach, using min-
imal syntax characters (#, *, -) to denote structure
rather than explicit opening/closing tags or key-value
pairs.Markdownprovidesagoodbalance:itachieves
native token efficiency without sacrificing immediate
human verification.
Table 2 presents examples of Markdown, JSON, and
XML formats. We compared these formats for the con-
strained environment of agricultural machinery.
5.5. Cross-lingual Tests
We assessed the models’ ability to retrieve informa-
tion across languages by posing English questions against
manuals in German, French, Dutch, Italian, and Spanish.
This test simulates a scenario where a user queries the
system in their native language (English in this test case)
while the machine’s internal documentation is stored in the
manufacturer’s original language or another available trans-
lation. We deliberately selected this cross-lingual retrieval
task(EnglishQuery→SourceLanguageDocument)toiso-
late cross-lingual question answering, following established
benchmarks like XTREME (Hu et al., 2020).
6. Results
6.1. Semantic Synchronization
The prototypeimplementation onthe physicalCAN bus
of ISO 11783 ETP transfer revealed significant differences
in transmission latency. We tested all combinations of six
languages (Dutch, French, English, German, Italian, and
Spanish) and three text structural encodings (Markdown,
XML and JSON), plus a few more variants of the ETP
parameter CTS. See Table 5 in Appendix D for detailed
results about transfer times and used data packets on the
CAN bus.
Summary of experimental results:
•Markdown:Markdown produced the smallest files,
leadingtoatransfertimeof47s(atCTS32;Figure3).
•XML:XML introduced approximately 5% overhead
compared to Markdown.
•JSON:The indented JSON files caused the highest
overhead(75%increase);withoutindentation,theEn-
glish JSON file would be 351166 bytes instead of
556063.
The transmission of visual data of graphics typical for
this operator manual was also tested (Figure 2). For the
data exchange experiment, the run-length encoded bitmap
T. Oksanen and J. Gun:Preprint Page 6 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
Table 2
Example of different formatting of text on page 2.
Markdown JSON XML

Machine identification

In order to provide you with good service various details
about your machine may sometimes be required.
Please complete these details here in order to
have them at hand when they are requested.
{ "page_number": 2,
"items": [ {
"type": "SectionHeaderItem",
"text": "Machine␣identification" }, {
"type": "TextItem",
"text": "In␣order␣to␣provide␣you␣with␣good␣service␣
various␣details␣about␣your␣machine␣may␣
sometimes␣be␣required.␣Please␣complete␣these
␣details␣here␣in␣order␣to␣have␣them␣at␣hand␣
when␣they␣are␣requested." },
<page number="2">
<text>Machine identification</text>
<text>In order to provide you with good service
various details about your machine may
sometimes be required. Please complete these
details here in order to have them at hand
when they are requested.</text>
(39855 bytes), which is defined in ISO 11783-6, was used.
Transferring these graphics with this encoding took 5.8 s.
The transfer times were measured at a background bus
load below 2%, which corresponds to the first coupling of
an implement (Section 5.1).
Figure 3:Comparison of measured data transfer time over
ISOBUS for different manual formats for the English manual
(CTS 32).
6.2. Formatting Efficiency
Formatchoiceaffectstransmissiontimeandcontextwin-
dow usage. Markdown was the most token-efficient format
for ingestion. The English version required approximately
30%fewertokensthantheotherEuropeanlanguages(Figure
4). This difference results from the tokenizer, which splits
non-English text into more tokens; the word counts of the
editions are similar (Section 4.2). Formatting choice sig-
nificantly affects token count. The JSON format adds sub-
stantial overhead, increasing token usage by approximately
95% compared to Markdown in the English dataset. We
note that token efficiency varies by model architecture and
tokenizer implementation; models optimized for non-Latin
scripts may exhibit different compression ratios.
6.3. Performance in LLM Tasks
We report model performance as F1 score; accuracy,
precision,recallandspecificitycanbederivedfromtheout-
comecategoriesinTable4.Inourevaluation,apositivecase
corresponds to an answerable question, and anegative case
corresponds to an unanswerable question. This distinction
Figure 4:Token efficiency comparison: XML vs JSON vs
Markdown across different languages.
allowsustoassessnotonlycorrectnessbutalsotheabilityof
the model to avoid hallucination (measured via specificity).
Asshownintheglobaloverview(Figure6),performance
varied by model size and architecture.
Figure 5 shows the F1 score of each model for the three
formats, a) for the English edition and b) for the five other
editionspooled.FortheEnglishedition,theformatsdiffered
by at most 0.05 for six of the eight models and by up to
0.12forthetwoDeepSeek-R1models,andnoneofthesedif-
ferences was significant. For the other editions, Markdown
exceeded JSON for Llama 3.1 8B (+0.09) and Qwen 2.5
7B (+0.08), whereas all other format differences were not
significant; confidence intervals are given in Appendix C.
FortheEnglisheditionwithMarkdown,F1rangedfrom
0.54 (DeepSeek-R1 1.5B) to 0.83 (Qwen 2.5 7B); the seven
models from 3B parameters upward reached 0.77 to 0.83
with overlapping confidence intervals, and GPT-OSS 120B
(0.81) did not exceed the 7B and 8B models (Appendix C).
6.4. Feasibility for Cross-lingual Tasks
The models demonstrated varying capabilities in cross-
lingualretrieval.AlthoughEnglishperformancewashighest
acrossalltestedmodels,weobservedadropinperformance
when querying manuals in other languages. With Mark-
down, F1 dropped by 0.12 to 0.30 from the English edition
to the other editions pooled; the drop was largest for Llama
3.2 3B (from 0.77 to 0.47), and for the other editions GPT-
OSS120B(0.68)andQwen38B(0.66)reachedthehighest
scores. Figure 7 confirms that smaller models performed
T. Oksanen and J. Gun:Preprint Page 7 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
Figure 5:F1 score by model and format: a) English edition, b) five other editions pooled.
Figure 6:F1 score by model and language edition for the three formats.
worse when the query language differs from the document
language.
Of all answers, 884 (5.7%) were written in the language
of the manual edition instead of English, most often by
Llama3.18B(248)andleastoftenbyGPT-OSS120B(31);
these answers were counted as incorrect (Appendix C).
The retrieval ablation with four models (Appendix E)
showed that, without retrieved context, three of the four
models declined every question. For the English edition,
BM25, semantic and hybrid retrieval did not differ signifi-
cantly. For the other editions, BM25 placed the gold page
(the page that has the correct answer) among the three re-
trievedchunksforonly9to24%ofthequestions,compared
with80to93%forsemanticretrieval,anditsF1was0.35to
0.47 lower than that of semantic and hybrid retrieval. With
the gold page given as context (oracle), F1 for the other
editions was 0.06 to 0.11 higher than with hybrid retrieval
(significant for 2/4 models), and the gap to the English
edition narrowed but remained (0.11 to 0.23).
7. Discussion
In this section, we interpret the experimental results to
define the architectural requirements for offline AI assis-
tants in agricultural machinery. We analyze the trade-offs
between network bandwidth and computational constraints.
Finally, we propose a verification framework to integrate
these stochastic models into safety-critical environments
while respecting the constraints of the ISO 11783 standard.
7.1. Protocol and Formatting
The results confirm that transferring operator manuals
over the CAN bus is feasible, provided that the system
uses an efficient format. We focus on the standard CAN
bus to ensure compatibility with the existing vehicle fleet.
The bandwidth constraints of ISO 11783 systems make
verbose object models (like hierarchical, indented JSON)
less attractive. Our results demonstrate that Markdown is
highlyeffective,offeringhightransferspeeds(47svs83sfor
JSON) by minimizing structural metadata while preserving
semantic hierarchy. Answer quality did not depend on the
T. Oksanen and J. Gun:Preprint Page 8 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
Figure 7:F1 score for the English edition and for the five other editions pooled, by model and format; points below the diagonal
indicate lower scores for the other editions.
format for most models (Section 6.3), so the choice of
Markdown rests on its size, transfer time and readability.
JSON had the lowest F1 for most models, significantly for
two.TheMarkdownmanuals(317to372kB)areofthesame
order of magnitude as the Virtual Terminal object pools
of commercial implements, as observed by the authors in
practice, and their transfer times (47 to 55 s at CTS 32) are
comparable to typical object pool uploads.
While Markdown minimizes transmission latency,
it shifts complexity to the receiving parser. Unlike
XML/JSON,whichhavemature,MISRA-compliantparsers,
Markdownparsingonsafety-criticalECUsrequiresrigorous
validationtoensurethereversetextmappingdoesnotdriftor
hallucinate the document structure. We accept this software
complexity trade-off to satisfy the bandwidth constraints of
theCANbus.Furthermore,thepresentationlayerassumesa
Next-GenVirtualTerminaloradedicatedtabletapplication,
as current ISO 11783-6 object pools possess limited font
support and cannot easily render the Unicode characters
required for multilingual Markdown display. The additional
space required in the Working Set ECU is reasonable for
text only, but the number of graphics increases the memory
requirements.
Faster ISOBUS leveraging Ethernet has been proposed.
A short-term solution proposal ISO FastLane (Oksanen,
2026) suggests that with a dual-stack approach, it is pos-
sible to accelerate transfer speeds significantly. While that
work focused on faster object pool uploads and faster Task
Controller rates, it could also accelerate transfer times for
operator manuals.
For standardization, the exchange tested here requires
only a new parameter group: the manual is requested with
thestandardRequestmessageandtransferredwithETP,both
of which are part of every ISO 11783 stack, and the version
check can follow the labels already used to identify Virtual
Terminal object pools and Task Controller device descrip-
tions.Becausethetestreliesonastandardstack—including
addressing, network management, request handling, and the
extendedtransportprotocol—implementingthisapplication
for any existing ISO 11783-compliant system is a sim-
ple addition to standard features. A standardized, machine-
readablemanualontheimplementwouldalsoallowcompa-
niesotherthantheimplementmanufacturertobuildproducts
onit,suchasassistants,diagnostictoolsortrainingapplica-
tions.
7.2. Initialization and Resource Management
The transfer times were measured under the conditions
of the first coupling of an implement (Section 5.1), when
the system is stationary, as for object pool uploads, and no
other large transfer runs at the same time. To keep these
conditions, we propose that the Semantic Synchronization
protocolberestrictedtoasafe-stateinitializationphase.The
system should initiate the transfer only when the machine
is in a Key-On/Engine-Off state or when the parking brake
is engaged. This protocol constraint ensures that the full
bandwidth is available for the manual transfer without com-
promising the safety-critical real-time control loops of the
machinery.
While the protocol solves the bandwidth issue, the sub-
sequent vector embedding process represents a significant
computational cost for the ECU. We acknowledge that real-
time embedding on an embedded CPU is unfeasible during
operation. With the first-coupling protocol and the version
check (Section 5.1), the manual is also embedded only
once, and the tractor caches the vector store in non-volatile
memory, which avoids both transfer and embedding for
daily operations. A CRC32 checksum of the manual is a
lightweight way to implement this version check; it detects
changes of the manual but does not authenticate it.
We addressed the trade-off between compression and
compute. Although applying compression algorithms (e.g.,
GZIP) would theoretically equalize the file sizes of JSON
and Markdown, we argue against their use in this specific
embedded context. Decompressing a large text corpus may
requiresignificantCPUcycles.InaKey-Onstartupscenario,
T. Oksanen and J. Gun:Preprint Page 9 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
the ECU’s CPU is already under heavy load initializing
real-time control loops. A raw text format like Markdown
avoids this decompression penalty, prioritizing immediate
availability over theoretical storage efficiency.
Ouranalysisrevealsauniqueopportunitytooptimizethe
system initialization by exploiting the decoupling of com-
putational and network resources. The initialization process
consists of two distinct phases: Data Transfer (Network-
bound: CAN Bus load) and Vector Embedding (Compute-
bound:CPU/NPUload).Becausethevectorembeddingpro-
cessonembeddedhardwareiscomputationallyintensiveand
time-consuming, it creates a window of significant duration
wheretheCANbusiseffectivelyidleregardingmanualdata.
We therefore propose an opportunistic background transfer
ofimages.Immediatelyafterreceivingthetext(Markdown),
the ECU begins the embedding process. Simultaneously,
the system initiates a low-priority background transfer of
the image library. Because the embedding process does not
consumenetworkbandwidth,thesetwotasksruninparallel.
Bythetimetheembeddingphaseiscompleteandthesystem
is ready to answer user queries, the system will have trans-
ferred a large portion (or all) of the image library. An on-
demandtransferofsingleimages(5.8sforthesampleimage,
Section 6.1) then serves only as a fallback mechanism for
anyimagesthatwerenotyetcapturedduringthebackground
phase. This approach is intended to avoid the startup delay
for text and to reduce the loading latency for images.
Weclarifythatthisarchitectureimplementslogicalmul-
timodality rather than visual processing. The 8B-parameter
models lack the computational capacity for runtime Vision
Transformer(ViT)inference.Therefore,thesystemretrieves
the Markdown image reference (e.g.,) based
onthesemanticsimilarityofthesurroundingtextorexplicit
imagedescriptions.Thisapproachavoidsthehighcomputa-
tionalcostofrunningvisiontransformersontheECUwhile
allowing the verification interface to display visual safety
data.
7.3. Context Preservation
Because standard RAG chunking strategies (e.g., 200-
token sliding windows) often sever the text from its par-
ent section headers, retrieving a chunk alone fails to pro-
vide the operator with the necessary document context. The
proposed self-describing architecture, however, solves this
efficiently through reverse text mapping. To generate the
UI breadcrumbs shown in our conceptual mockup (Figure
8), the system would perform a recursive backward search
from the matched chunk to identify the preceding H1, H2,
and H3 headers within the Markdown source. Markdown
allows for header-based anchoring with zero database over-
head.Becausethetractorterminalholdsthecompletesource
Markdown file in memory, it can perform an exact string
searchtolocatetheretrievedchunkwithinthefulldocument
(Appendix E). This operation is computationally negligible
(microseconds) for manual-sized documents.
This capability highlights a secondary advantage of
Markdown over JSON: Markdown preserves the linear,
human-readable text stream. To display the context in a
JSON-basedsystem,theterminalwouldneedtoreconstruct
the document flow from the hierarchical object tree. With
Markdown, the terminal simply renders the file, performs
a string match for the retrieved answer, scrolls to that
location, and highlights it. This feature allows the operator
to instantly verify the AI’s advice against the surrounding
safety warnings and diagrams.
7.4. Hardware Outlook
We acknowledge that deployment on legacy automotive
ECUsisunfeasibleonthetractorside,whichinourconcept
is the consumer of the operator manual data, while the
tested concept should work directly for legacy ECUs on the
implementside.However,theemergenceofnext-generation
automotive compute platforms could render this architec-
ture viable. The assumption that agricultural vehicles have
available embedded nodes is realistic. ECU suppliers have
recently introduced high-performance computers (HPCs) in
industriallymanufacturedvehicleplatforms.Platformssuch
as the NVIDIA DRIVE Thor and the Qualcomm Snap-
dragon Cockpit Elite both feature the native 4-bit hardware
acceleration and high-bandwidth unified memory architec-
tures required to run the models we tested locally.
7.5. Safety and Verification
To enable the reverse text mapping shown in Figure
8, the receiving ECU must implement a robust Markdown
syntax tree parser to reliably identify parent headers from
retrievedtextchunks,whereasJSONprovidesthishierarchy
explicitly.Although8B-parametermodelsdemonstratehigh
retrieval accuracy, the risk of hallucination—particularly
regardingnumericalsafetydataliketorquesettings—cannot
be eliminated. Furthermore, deploying 4-bit quantized
weights on edge processors introduces known risks of
numerical degradation. Quantization clipping can silently
corrupt specific float values generated by the model,
potentially altering a 50 Nm torque specification to 5
Nm (Gong et al., 2025). This hardware-level limitation
elevatestheproposedverificationinterfacefromaninterface
convenience to a critical safety requirement. The system
relies entirely on the exact-string reverse mapping of the
source Markdown to present the original numerical data
alongsidethegeneratedtext.Weacknowledgethatproviding
a generated answer introduces the risk of automation
bias, where operators may overlook the source text. The
verification interface is intended to reduce this risk, not to
remove it.
Therefore, we propose that an agricultural RAG system
support verification of its answers: the system guides the
operator to the source text instead of acting as the final
authority. As proposed in Figure 8, the system should not
present the LLM’s output as an authoritative end-state. In-
stead, theVerify Sourcetrigger opens a split-screen view,
aligning the generative response with the primary safety
documentation.ThischangeshiftstheroleoftheLLMfrom
an oracle to a navigator, so that the operator can visually
verify the AI’s advice against the displayed source text.
T. Oksanen and J. Gun:Preprint Page 10 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
Operator manual assistant Manual of the connected implement 14:32
Operator manual (English) › Page 43 › Vanes Source 1 / 3
The spreader is provided with 12 vanes of equal length and 4 shorter.
Grease the threaded rods of the vanes.
Hook the short vanes into the correct position (see illustration).
Then place the other two vanes into the other positions.
Press the vanes in well and put the lock nut in position.
Place the vanes on the other side in the same manner.
Fasten the lock nuts (50 Nm)
Even wear of the vanes
The long vanes are positioned in groups of three.
The first vane considered in the direction of rotation of the disc will wear out
most quickly.
TIP Exchange the first vane after the fixed vane, and the short vane, over a
period of time.
After a subsequent period, exchange the foremost vane at that time with
the central vane.
Today, 14:31
What torque (in Nm) should be
applied to the vane lock nuts?
Operator
The vane lock nuts should be fastened to
50 Nm.
Verify: please check the highlighted passage in
the manual.
Verify source
Ask the manual…
Figure 8:Conceptual mockup of the proposed verification interface. The interface displays the generated answer (left) alongside
the original source text (right) for verification by the operator.
Rather than acting as an authoritative oracle, the proposed
interfaceimplementsahuman-in-the-loopverificationlayer,
a design pattern intended to maintain the operator’s role as
the primary decision-maker, consistent with safety princi-
ples for information systems.
We acknowledge a residual risk regarding visual
salience. If the retriever highlights an incorrect section
(e.g., tire pressure instead of torque), the operator might
succumb to automation bias and trust the highlighted
sectionwithoutreadingthetext.Thearchitectureprioritizes
the verbatim rendering of the source Markdown, so that the
operator can base the final decision on the source text.
We argue that the ISOBUS network is a medium that
keeps the operator manual physically attached to the ma-
chinethroughoutitslifecycle.Unlikecloudsolutions,which
dependonconnectivity,orUSBsticks,whichareeasilylost,
the ECU-stored manual travels with the steel. This local
storage ties the documentation to the specific serial number
configuration and prevents the drift often observed in cloud
systemswheremanualsdocumentnewerhardwarerevisions
than the physical machine possesses. Consequently, the op-
erator has the documentation that matches the machine,
regardless of ownership changes.
7.6. Model Capabilities
Ifamodelcannotperformthetaskonhigh-performance
hardware, it will fail on an embedded ECU. Although this
study uses PC-grade inference, current ISO 11783-6 Vir-
tual Terminals typically lack dedicated NPU acceleration.
Therefore, we propose that this architecture requires a ded-
icated AI compute node on the CAN bus, separate from
the standard Virtual Terminal. The 8B-parameter models
would require 4-bit quantization (e.g., GGUF/EXL2) and
approximately 5 to 8 GB of VRAM.
The performance differences between LLMs highlight
that the 120B-parameter model offered no advantage over
7Band8BmodelssuchasQwen2.5andQwen3inthistask
(Section 6.3), which makes smaller models candidates for
running these assistants on edge devices within the tractor
cabin. Users must, however, exercise caution. As noted in
our previous work (Gun and Oksanen, 2025), hallucination
remains a risk,particularly for unanswerable questions.The
number of hallucinations of each model is given in Table 4.
Cross-lingual capabilities are promising but imperfect
with the tested LLMs. For safety-critical applications,
native-language manuals are preferable to relying solely
on the translation abilities of the LLM. Although high-
performing models can bridge the language gap effectively,
thedropinperformanceweobservedindicatesthataccuracy
degrades compared to English questions on the English
edition. The retrieval ablation shows that semantic retrieval
was essential for the other editions, because keyword
retrieval failed with English questions on the translated
text (Section 6.4). Even with the gold page as context, a
gap of 0.11 to 0.23 to the English edition remained, so the
T. Oksanen and J. Gun:Preprint Page 11 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
remaining drop is not caused by retrieval alone. Part of it
came from answers written in the language of the manual
instead of English, which also occurred with the gold page
(22 to 53 of 540 answers per model).
7.7. Limitations
The study used the operator manual of one fertilizer
spreader. The six language editions are translations of the
same manual; they test cross-lingual behaviour but are not
independentvalidationsets.The108questionswerewritten
by the authors, their balance of answerable and unanswer-
able questions does not reflect the questions of operators
in practice, and the same questions were asked for every
edition, so the observations are correlated. All questions
wereposedinEnglish,sonative-languageusewasnottested.
The answers were judged automatically. The judge
agreed with four human raters for 93.5% of a stratified
sample (Appendix F), but its verdicts on individual answers
can still be wrong. Each configuration was run once at
temperature0;repeatedrunsandothersoftwareorhardware
versions can produce different outputs (up to 0.04 F1 in
this study, Appendix E). The retrieval ablation covered four
models, the Markdown editions and one set of chunking
parameters (200 tokens, 100 tokens overlap, three chunks)
and all models used the same embedding model.
Inference ran on PC-grade hardware. Latency, memory
use, energy consumption and real-time behaviour on an
ECU suitable for the tractor cabin were not tested, nor was
the effect of 4-bit quantization on retrieval accuracy and
numerical values. The transfers were measured once per
configuration on a laboratory bus at the low background
load of the first coupling of an implement. Images were not
retrieved by their content.
The results refer to the tested serialization (indented
JSON,XMLwithoutitemtypes,Markdown);minified,com-
pressed or binary representations were not evaluated. The
PDF manuals were converted with a wrapper using Docling
and used as converted. The proposed verification interface
was not evaluated with operators, the derivation of the
headingpathfromalocatedchunkwasnottested,thefidelity
of the PDF conversion was not quantified, and integrity and
security aspects such as manipulated manuals or prompt
injection were outside the scope of this study.
8. Conclusions
In this study, it was demonstrated that a complete op-
erator manual can be transferred over ISO 11783 with the
standard request message and the extended transport pro-
tocol in 31 to 89 s, which is within the time of a typical
object pool upload at the first coupling of an implement.
Among the three formats, Markdown produced the smallest
files,theshortesttransfertimesandthefewesttokens,while
answer quality was comparable across the formats for most
models. Locally run open-weight LLMs answered opera-
tor questions from the transferred manual: for the English
edition, the models from 3B parameters upward reached
F1 scores of 0.77 to 0.83, without an advantage of the
120B model, and for the other editions the scores were
0.12 to 0.30 lower. Semantic retrieval was essential across
languages,becausekeywordretrievalfailedforthetranslated
editions,andthehybridcombinationgavenoadvantageover
semanticretrievalalone;evenwiththegoldpage,partofthe
drop remained, partly because the models answered in the
language of the manual.
Two design elements were also proposed but not eval-
uated in this study: a first-coupling protocol with a version
check, so that the manual is transferred and embedded only
once per implement revision, and an on-demand or back-
ground transfer of images, for which a single compressed
imagewastransferredin5.8s.Thenextstepisabenchmark
of the assistants on ECU hardware suitable for the tractor
cabin.
Glossary
CANController Area Network
ECUElectronic Control Unit
ETPExtended Transport Protocol
GNSSGlobal Navigation Satellite System
ISOBUSISO 11783 standard
LLMLarge Language Model
PGNParameter Group Number
RAGRetrieval-Augmented Generation
Acknowledgements
The authors thank the Kverneland Group for providing
the open-access manuals used in this study. We thank espe-
cially Peter van der Vlugt for his support to get insight for
these publicly available operator manual documents.
Data Availability Statement
The ManualSync code and the question-answer dataset
are available athttps://github.com/TUMAMX/agriquery/tree/
main/ManualSync.
Declaration of generative AI and AI-assisted
technologies in the manuscript preparation
process
During the preparation of this work the authors used
Gemini 3 Pro Preview and Gemini 3 Flash Preview to
paraphrase the manuscript text, and Claude (Anthropic) to
assist in improving the text and the code. After using these
tools,theauthorsreviewedandeditedthecontentasneeded
and take full responsibility for the content of the published
article.
T. Oksanen and J. Gun:Preprint Page 12 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
References
, 2013. ISO 4254-1:2013 Agricultural machinery — Safety — Part 1:
Generalrequirements. URL:https://www.iso.org/standard/55910.html.
confirmed 2018.
, 2015. ISO 11783-10:2015 Tractors and machinery for agriculture and
forestry — Serial control and communications data network — Part 10:
Task controller and management information system data interchange.
URL:https://www.iso.org/standard/61581.html.
, 2017. ISO 11783-1:2017 Tractors and machinery for agriculture and
forestry — Serial control and communications data network — Part 1:
General standard for mobile data communication. URL:https://www.
iso.org/standard/57556.html.
, 2018a. ISO 11783-3:2018 Tractors and machinery for agriculture and
forestry — Serial control and communications data network — Part 3:
Data link layer. URL:https://www.iso.org/standard/71172.html.
, 2018b. ISO 11783-6:2018 Tractors and machinery for agriculture and
forestry — Serial control and communications data network — Part 6:
Virtual terminal. URL:https://www.iso.org/standard/71173.html.
,2018c.ISO4254-8:2018Agriculturalmachinery—Safety—Part8:Solid
fertilizer distributors.
, 2019a. ISO 11783-12:2019 Tractors and machinery for agriculture and
forestry — Serial control and communications data network — Part 12:
Diagnostics services. URL:https://www.iso.org/standard/71184.html.
, 2019b. ISO 11783-2:2019 Tractors and machinery for agriculture and
forestry — Serial control and communications data network — Part 2:
Physical layer.
Akbar, N.A., Dembani, R., Lenzitti, B., Tegolo, D., 2025. RAG-driven
memoryarchitecturesinconversationalLLMs—aliteraturereviewwith
insights into emerging agriculture data sharing. IEEE Access 13,
123855–123880. doi:10.1109/ACCESS.2025.3589241.
Ateş, Y., Sayar, A., Bozlar, İ.U., Ertuğrul, S., Arslan, S.S., 2025. Semantic
chunkingandchain-of-thoughtreasoningforRAG-baseddocumentpro-
cessing,in:2025IEEE35thInternationalWorkshoponMachineLearn-
ingforSignalProcessing(MLSP),pp.1–6. doi:10.1109/MLSP62443.2025.
11204203.
Auer, C., Lysak, M., Nassar, A., Dolfi, M., Livathinos, N., Vagenas, P.,
Ramis, C.B., Omenetti, M., Lindlbauer, F., Dinkla, K., Mishra, L.,
Kim, Y., Gupta, S., de Lima, R.T., Weber, V., Morin, L., Meijer, I.,
Kuropiatnyk, V., Staar, P.W.J., 2024. Docling technical report. URL:
https://arxiv.org/abs/2408.09869,arXiv:2408.09869.
Bettucci, F., Sozzi, M., Benetti, M., Sartori, L., 2024. A data-
driven approach to agricultural machinery working states analysis
during ploughing operations. Smart Agricultural Technology 8,
100511. URL:https://www.sciencedirect.com/science/article/pii/
S2772375524001163, doi:https://doi.org/10.1016/j.atech.2024.100511.
Chen, Y., Li, C., Yuan, Q., Li, J., Fan, Y., Ge, X., Li, Y., Gao, F., Zhao,
R., 2025. Cockpit-Llama: Driver intent prediction in intelligent cockpit
via large language model. Sensors 25. URL:https://www.mdpi.com/
1424-8220/25/1/64, doi:10.3390/s25010064.
Chua, L., Ghazi, B., Huang, Y., Kamath, P., Kumar, R., Manurangsi, P.,
Sinha, A., Xie, C., Zhang, C., 2025. Crosslingual capabilities and
knowledgebarriersinmultilinguallargelanguagemodels. URL:https:
//arxiv.org/abs/2406.16135,arXiv:2406.16135. updated March 2025.
Cormack, G.V., Clarke, C.L.A., Buettcher, S., 2009. Reciprocal rank
fusion outperforms Condorcet and individual rank learning methods,
in: Proceedings of the 32nd International ACM SIGIR Conference on
Research and Development in Information Retrieval, Association for
ComputingMachinery,NewYork,NY,USA.pp.758–759. URL:https:
//doi.org/10.1145/1571941.1572114, doi:10.1145/1571941.1572114.
EU, 2006. Directive 2006/42/EC of the European Parliament and of the
Councilof17May2006onmachinery. OfficialJournaloftheEuropean
Union. URL:http://data.europa.eu/eli/dir/2006/42/oj.
Gong, R., Ding, Y., Wang, Z., Lv, C., Zheng, X., Du, J., Qin, H., Guo, J.,
Magno, M., Liu, X., 2025. A survey of low-bit large language models:
Basics, systems, and algorithms. URL:https://arxiv.org/abs/2409.
16694,arXiv:2409.16694.
Gun, J., Oksanen, T., 2025. Agri-Query: A case study on RAG vs. long-
context LLMs for cross-lingual technical question answering. URL:
https://arxiv.org/abs/2508.18093,arXiv:2508.18093.
Hu, J., Ruder, S., Siddhant, A., Neubig, G., Firat, O., Johnson, M., 2020.
XTREME: A massively multilingual multi-task benchmark for evalu-
ating cross-lingual generalization. URL:https://arxiv.org/abs/2003.
11080,arXiv:2003.11080.
Jafarnejad, S., Berthe-Pardo, A., Frank, R., 2024. Demo: Towards a con-
versational LLM-based voice assistant for transportation applications,
in: 2024 IEEE Vehicular Networking Conference (VNC), pp. 261–
262. URL:https://ieeexplore.ieee.org/document/10575993, doi:10.
1109/VNC61989.2024.10575993.
Johnson, E., Wilson, N., 2024. Enhancing agricultural machinery
management through advanced LLM integration. arXiv preprint
arXiv:2407.20588 URL:https://arxiv.org/abs/2407.20588.
Kaur, R., Bhankhar, A.S., Salh, J.S., Rajput, S., Vidhi, Mahendra, K.,
Berwal, B., Kumar, R., Ranathunga, S., 2025. Leveraging synthetic
data for question answering with multilingual LLMs in the agricultural
domain. URL:https://arxiv.org/abs/2507.16974,arXiv:2507.16974.
Koopman, B., Mourad, A., Li, H., van der Vegt, A., Zhuang, S., Gibson,
S., Dang, Y., Lawrence, D., Zuccon, G., 2022. AgAsk: An agent to
helpanswerfarmer’squestionsfromscientificdocuments. URL:https:
//arxiv.org/abs/2212.10762,arXiv:2212.10762.
Kverneland Group, 2021a. Gebruikershandleiding - Exacta-TLX-
GEOSPREADGS3. KvernelandGroupNieuw-VennepB.V..Hoofdweg
1278, NL-2153 LR Nieuw-Vennep, The Netherlands. URL:https://
www.kvgportal.com/W_global/Media/lexcom/VN/A14870/A148703740-2.pdf.
reference Number: A148703740-2; Language: Dutch.
Kverneland Group, 2021b. Manual de usuario - Exacta-TLX-
GEOSPREADGS3. KvernelandGroupNieuw-VennepB.V..Hoofdweg
1278, NL-2153 LR Nieuw-Vennep, The Netherlands. URL:https:
//www.kvgportal.com/W_global/Media/lexcom/VN/A14880/A148818340.pdf.
reference Number: A148818340; Language: Spanish.
Kverneland Group, 2021c. Manuale di istruzioni - Exacta-TLX-
GEOSPREADGS3. KvernelandGroupNieuw-VennepB.V..Hoofdweg
1278, NL-2153 LR Nieuw-Vennep, The Netherlands. URL:https:
//www.kvgportal.com/W_global/Media/lexcom/VN/A14880/A148818540.pdf.
reference Number: A148818540; Language: Italian.
Kverneland Group, 2021d. Manuel d’utilisation - Exacta-TLX-
GEOSPREADGS3. KvernelandGroupNieuw-VennepB.V..Hoofdweg
1278, NL-2153 LR Nieuw-Vennep, The Netherlands. URL:https://
www.kvgportal.com/W_global/Media/lexcom/VN/A14870/A148703640-2.pdf.
reference Number: A148703640-2; Language: French.
KvernelandGroup,2021e.UserManual-Exacta-TLX-GEOSPREADGS3.
Kverneland Group Nieuw-Vennep B.V.. Hoofdweg 1278, NL-2153 LR
Nieuw-Vennep, The Netherlands. URL:https://www.kvgportal.com/
W_global/Media/lexcom/VN/A14870/A148703540-2.pdf. reference Number:
A148703540-2; Language: English; Applicable from software version
V: GS3 x.x., machine number 1001, serial number VN407.
Kverneland Group, 2024. Betriebsanleitung - Exacta-TLX-GEOSPREAD
GS3. Kverneland Group Nieuw-Vennep B.V.. Hoofdweg 1278,
NL-2153 LR Nieuw-Vennep, The Netherlands. URL:https://www.
kvgportal.com/W_global/Media/lexcom/VN/A14880/A148818240-1.pdf. ref-
erence Number: A148818240-1; Language: German.
Li, Y., et al., 2024. Large language models for human-like autonomous
driving: A survey. arXiv preprint arXiv:2407.19280 URL:https://
arxiv.org/abs/2407.19280.
Liu, N.F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F.,
Liang, P., 2023. Lost in the middle: How language models use long
contexts. URL:https://arxiv.org/abs/2307.03172,arXiv:2307.03172.
Oksanen, T., 2026. ISO FastLane: Faster ISO 11783 with dual stack
approach as a short term solution. URL:https://arxiv.org/abs/2602.
09633,arXiv:2602.09633.
Rohmadin, N.A., Ferdiana, R., Hidayah, I., 2025. Optimizing retrieval-
augmentedgenerationchatbotwithhyperparametertuning,in:20254th
International Conference on Electronics Representation and Algorithm
(ICERA), pp. 623–628. doi:10.1109/ICERA66156.2025.11087302.
Saha, B., Saha, U., Malik, M.Z., 2024. QuIM-RAG: Advancing retrieval-
augmentedgenerationwithinvertedquestionmatchingforenhancedQA
performance. IEEE Access 12, 185401–185410. doi:10.1109/ACCESS.
T. Oksanen and J. Gun:Preprint Page 13 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
2024.3513155.
Samuel, D.J., Skarga-Bandurova, I., Sikolia, D., Awais, M., 2025.
AgroLLM: Connecting farmers and agricultural practices through large
languagemodelsforenhancedknowledgetransferandpracticalapplica-
tion. URL:https://arxiv.org/abs/2503.04788,arXiv:2503.04788.
da Silva, C.A.G., Rodrigues de Sá, J.L., Menegatti, R., 2019. Diagnostic
of failure in transmission system of agriculture tractors using predictive
maintenancebasedsoftware. AgriEngineering1,132–144. URL:https:
//www.mdpi.com/2624-7402/1/1/10, doi:10.3390/agriengineering1010010.
Sivakumar, N., Srinivasan, P., Kannan, M., Vijayalakshmi, P., T,
N., 2023. PooRaa-Agri KG: An agricultural knowledge graph-
based simplified multilingual query system. Expert Systems
40. URL:https://onlinelibrary.wiley.com/doi/abs/10.1111/exsy.
13434, doi:10.1111/exsy.13434.
Vizniuk, A., Diachenko, G., Laktionov, I., Siwocha, A., Xiao, M., Smoląg,
J., 2025. A comprehensive survey of retrieval-augmented large lan-
guage models for decision making in agriculture: unsolved problems
and research opportunities. Journal of Artificial Intelligence and Soft
Computing Research 15, 115–146. doi:10.2478/jaiscr-2025-0007.
Wang, L., Jin, T., Yang, J., Leonardis, A., Wang, F., Zheng, F., 2024.
Agri-LLaVA: Knowledge-infused large multimodal assistant on agri-
cultural pests and diseases. URL:https://arxiv.org/abs/2412.02158,
arXiv:2412.02158.
Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y.,
Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.,
2023. Judging LLM-as-a-judge with MT-bench and Chatbot Arena, in:
Proceedingsofthe37thInternationalConferenceonNeuralInformation
Processing Systems, Curran Associates Inc., Red Hook, NY, USA.
URL:https://proceedings.neurips.cc/paper_files/paper/2023/hash/
91f18a1287b398d378ef22505bf41832-Abstract-Datasets_and_Benchmarks.
html.
A. Appendix: QA Test Dataset
A.1. Answerable Question Samples
•Q:Whattorque(inNm)shouldbeappliedtothevane
lock nuts? A: 50 Nm
•Q: What is the required grease level in mm below the
filler opening for spreading disc gearboxes after the
machine has stood still? A: 35 mm
•Q:Whereisthemainswitchbuttontoturnthecontrol
box on or off located? A: The main switch button is
located on the upper left in the red extension.
•Q: How to enable fine application for dosing low
application rate? A: Move the fine application handle
to the fine dosing position on both sides.
•Q: Where is the RS 232 connector located? A: At the
back of the control box.
•Q: How often should the agitator axle seal be re-
placed? A: Every season and after every 100 opera-
tional hours.
•Q: From which machine point is the spreading height
measured to the ground or the crop? A: Measured
from the bottom of the vanes.
•Q: What materials are required to perform the tray
test? A: A measuring tape or ruler, a spirit level, 7
troughs, 7 graduated tubes, a funnel, a notebook, pen,
calculator, this manual, and the software’s instruction
manual.
•Q: How many parts does the distribution meter have?
A: Seven.
•Q:Shouldtheparkingbrakeofthetractorbeengaged
before connecting the machine? A: Yes.
•Q:Howlongshouldthemainswitchbuttonbepressed
toturnthecontrolboxonoroff?A:Atleast1second.
•Q: What is the overlap percentage for the full field
spreading pattern? A: 100% overlap.
•Q:Whatdeterminesthemachine’sworkingwidth?A:
Spreading disc RPM
•Q: When shortening a coupling shaft, how far must
profiled tubes at least overlap in mm? A: 150 mm.
A.2. Unanswerable Question Samples
•Q:Howmuchextradieseldoesthetractorconsumeto
use the Exacta-TLX GEOSPREAD? A: Not found in
context
•Q:Isone-sidedboundaryspreadingsuitableforsmall
gardens? A: Not found in context
•Q: Can IsoMatch Tellus be connected to an external
mouse? A: Not found in context
•Q: Can the linkage pin for the tractor be made out of
aluminium? A: Not found in context
•Q: What is the maximum height the fertilizer flies
when spreading without GEOCONTROL headland?
A: Not found in context
•Q: What happens if the machine grease nipples are
never lubricated? A: Not found in context
•Q: What kind of protective safety gloves are needed
for cleaning fertiliser remnants from the Exacta-TLX
GEOSPREAD before welding? A: Not found in con-
text
•Q: What specific ’grease’ type is recommended for
’profiled tubes’ of the coupling shaft? A: Not found
in context
•Q: What is the minimum ’baud rate’ for ’RS 232
connection’? A: Not found in context
•Q: What is the recommended tire pressure ’range’ for
transport mode? A: Not found in context
•Q: What is the drain rate in liters per minute of ’drain
kit’ for hopper emptying? A: Not found in context
•Q: How many hours of continuous operation can the
IsoMatch Tellus operate before flattening a typical
tractor battery if left switched on with the engine off?
A: Not found in context
T. Oksanen and J. Gun:Preprint Page 14 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
•Q: What specific paint should be used to paint any
damaged paintwork at the end of the season the ma-
chine in preparation for winter storage? A: Not found
in context
•Q: Can the IsoMatch universal ISOBUS terminal be
used to check the weather? A: Not found in context
B. Appendix: Prompts
B.1. Question prompts
<purpose>Extract a precise, concise answer to the question from the
given context. Adhere strictly to the instructions. Base your answer
on the context.</purpose>
<instructions>
<instruction>Read the entire context carefully</instruction>
<instruction>Focus ONLY on the specific information related to
the question</instruction>
<instruction>Provideanextremelypreciseanswer</instruction>
<instruction>Match the expected answer format exactly
</instruction>
<instruction>If unsure, respond with "Unknown" or "Not found
in context"</instruction>
<instruction>Answer in English</instruction>
</instructions>
<context>
context
</context>
<question>
question
</question>
<answer>
[Carefully extract the EXACT information that directly answers
the question, keeping it as brief and precise as possible]
</answer>
B.2. Evaluation Prompt
<purpose>ANSWER COMPARISON TASK. Do ANSWER_ONE
and ANSWER_TWO convey the same information regarding the
QUESTION? Adhere strictly to the INSTRUCTIONS. Base your
ANSWER on the CONTEXT.</purpose>
<INSTRUCTIONS>
<instruction>- Respond ’yes’ if ANSWER_ONE and the
ANSWER_TWO convey the SAME TECHNICAL MEANING
</instruction>
<instruction>- Consider ’yes’ if differences are
INSIGNIFICANT to the core technical content</instruction>
<instruction>- Respond ’no’ ONLY if there are MEANINGFUL
differences that alter the technical understanding</instruction>
<instruction>- Assess the SUBSTANCE of the information, not
surface-level variations</instruction>
<instruction>- Answer ONLY with yes or no</instruction>
<instruction>- Don’t provide additional information
</instruction>
</INSTRUCTIONS>
<CONTEXT>
<QUESTION>
question
</QUESTION>
<ANSWER_ONE>
model_answer
</ANSWER_ONE>
<ANSWER_TWO>
expected_answer
</ANSWER_TWO>
</CONTEXT>
<ANSWER>
(yes/no)
</ANSWER>
C. Appendix: Detailed Test Results
Table 3 reports the F1 score of every model for each
format separately for the English edition and for the five
other editions pooled with question-level bootstrap confi-
dence intervals and the scoring rule of Section 5.3.
Table 4 breaks down the answers of each model into
outcome categories, pooled over the six language editions
andthethreeformats.Foranswerablequestions,anincorrect
answer was classified as a false refusal, if it stated that the
information was not available; and otherwise as a wrong
answer. Answers written in the language of the operator
manual edition, instead of English, are listed separately
whatever the verdict of the judge.
Of all answers, 884 (5.7%) were written in the language
of the manual edition, all of them for the five non-English
editions. The LLM judge accepted 355 of them, because
their content matched the expected answer. Scoring these
answersbytheverdictofthejudge(insteadofcountingthem
as incorrect) would raise the pooled F1 score for the non-
English editions by 0.01 to 0.10 (depending on the model).
The results for the English edition are unaffected.
D. Appendix: Transfer Protocol and
Measurements
D.1. Data Transfer Performance Data
Table 5 presents the measurements from the ISO 11783
ETPtransfertests.ThetestswerecarriedoutwithtwoWin-
dows desktop PCs, each connected to the physical bus with
a Kvaser Leaf Light v2 USB-CAN interface and running an
in-house C implementation of addressing, request handling
and ETP on the Kvaser CANlib library. The bus had a 3
m backbone with stubs of at most 0.5 m and a terminating
bias circuit at each end, and it also carried an AEF-certified
commercial Virtual Terminal server, Task Controller server
and file server. Each combination was transferred once, and
transfer times were taken from server-side timestamps.
The results highlight the effect of the Clear to Send
(CTS)parameter;increasingCTSfrom16to128packetsper
burst for the Dutch manual reduced transfer time from 66.6
s to 30.8 s. The indented JSON files (Section 5.4) required
the longest transfer times. The compressed sample image
(image_test) was transferred in 5.8 s.
D.2. Format Efficiency
E. Appendix: Retrieval Ablation
The retrieval configuration of RAG was evaluated in
two parts: a) a retrieval-only evaluation without a language
T. Oksanen and J. Gun:Preprint Page 15 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
Table 3
F1 score [95% confidence interval] by model and format.
Model Format English edition Other editions (pooled)
DeepSeek-R1 1.5B Markdown 0.54 [0.45, 0.62] 0.37 [0.29, 0.45]
JSON 0.61 [0.51, 0.70] 0.32 [0.24, 0.39]
XML 0.66 [0.58, 0.73] 0.32 [0.24, 0.39]
Llama 3.2 3B Markdown 0.77 [0.68, 0.85] 0.47 [0.36, 0.57]
JSON 0.77 [0.69, 0.85] 0.46 [0.35, 0.56]
XML 0.72 [0.62, 0.80] 0.50 [0.39, 0.59]
Qwen 2.5 7B Markdown 0.83 [0.75, 0.90] 0.65 [0.56, 0.73]
JSON 0.80 [0.71, 0.87] 0.56 [0.47, 0.65]
XML 0.84 [0.77, 0.90] 0.61 [0.51, 0.70]
Llama 3.1 8B Markdown 0.78 [0.70, 0.85] 0.60 [0.51, 0.68]
JSON 0.76 [0.68, 0.84] 0.51 [0.42, 0.60]
XML 0.79 [0.71, 0.87] 0.53 [0.44, 0.62]
DeepSeek-R1 8B Markdown 0.81 [0.74, 0.88] 0.63 [0.55, 0.70]
JSON 0.78 [0.70, 0.86] 0.61 [0.54, 0.69]
XML 0.73 [0.64, 0.81] 0.58 [0.50, 0.66]
Qwen 3 8B Markdown 0.78 [0.70, 0.86] 0.66 [0.58, 0.74]
JSON 0.78 [0.70, 0.86] 0.66 [0.58, 0.74]
XML 0.80 [0.72, 0.88] 0.62 [0.54, 0.70]
Gemma 3 12B Markdown 0.78 [0.71, 0.85] 0.64 [0.56, 0.71]
JSON 0.74 [0.65, 0.82] 0.59 [0.50, 0.66]
XML 0.78 [0.70, 0.85] 0.60 [0.51, 0.68]
GPT-OSS 120B Markdown 0.81 [0.74, 0.89] 0.68 [0.60, 0.75]
JSON 0.78 [0.70, 0.85] 0.64 [0.56, 0.72]
XML 0.78 [0.70, 0.85] 0.66 [0.59, 0.73]
Table 4
Outcome categories per model (number of answers; 1944 per model, 972 each to answerable and unanswerable questions).
Answers in the language of the operator manual are counted for both question types.
Answerable questions Unanswerable questions
Model Correct
Wrong
answer
False
refusal
Language
of manual
Correct
abstention Hallucination
DeepSeek-R1 1.5B 367 420 161 45 410 541
Llama 3.2 3B 419 104 340 160 784 137
Qwen 2.5 7B 545 123 190 147 806 133
Llama 3.1 8B 521 158 118 248 702 197
DeepSeek-R1 8B 557 232 140 59 756 200
Qwen 3 8B 609 216 104 53 748 214
Gemma 3 12B 590 195 97 141 682 239
GPT-OSS 120B 649 211 90 31 691 272
Figure 9:Token and character counts of the converted manuals
byformatandlanguageedition(tokens:Qwen3-Embedding-8B
tokenizer).
model and b) an end-to-end ablation with five context con-
ditions. Our earlier study with the same question set com-
pared keyword, semantic and hybrid retrieval with another
embedding model (Gun and Oksanen, 2025); the hybrid
configuration, which performed best there, was retained for
the main experiments of this study.
E.1. Retrieval-only evaluation
Forthe54answerablequestions,thechunksofthemain
experiments were ranked by BM25; by semantic search and
by hybrid fusion (RRF with𝑘= 60and ten candidates per
retriever). A retrieved chunk counted as a hit, if it covered
the gold page of the question; each chunk was mapped to
itspagesthroughthepagemarkersoftheconvertedoperator
manual. For this mapping, each chunk was located by exact
string search in its converted manual; all chunks of the 18
collections were found verbatim. Table 6 reports Recall@3,
T. Oksanen and J. Gun:Preprint Page 16 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
Table 5
Detailed ETP data transfer results comparing languages, formats, and protocol settings (CTS). "DT" refers to Data Transfer
packets; "CM" refers to Connection Management packets.
Filename Size (B) DT Pkts CM Pkts Load Time Param
dutch_manual.md 319540 45649 1 + 2855 41.1% 66.6 s CTS 16
dutch_manual.md 319540 45649 1 + 1428 55.6% 47.1 s CTS 32
dutch_manual.md 319540 45649 1 + 715 70.9% 35.4 s CTS 64
dutch_manual.md 319540 45649 1 + 358 79.1% 30.8 s CTS 128
dutch_manual.xml 334470 47782 1 + 1495 55.5% 48.9 s CTS 32
dutch_manual.json 555565 79367 1 + 2482 54.8% 84.2 s CTS 32
english_manual.md 317454 45351 1 + 1419 55.3% 47.0 s CTS 32
english_manual.xml 333699 47672 1 + 1491 54.9% 49.3 s CTS 32
english_manual.json 556063 79438 1 + 2484 55.6% 82.7 s CTS 32
french_manual.md 371708 53102 1 + 1661 55.6% 54.9 s CTS 32
french_manual.xml 387402 55344 1 + 1731 55.4% 57.0 s CTS 32
french_manual.json 607023 86718 1 + 2711 56.8% 89.1 s CTS 32
german_manual.md 347323 49618 1 + 1552 55.9% 50.7 s CTS 32
german_manual.xml 363334 51905 1 + 1624 55.4% 53.1 s CTS 32
german_manual.json 585542 83649 1 + 2616 56.4% 86.4 s CTS 32
italian_manual.md 360976 51568 1 + 1613 55.4% 52.9 s CTS 32
italian_manual.xml 377485 53927 1 + 1687 55.0% 56.0 s CTS 32
italian_manual.json 597026 85290 1 + 2667 57.4% 86.3 s CTS 32
spanish_manual.md 358426 51204 1 + 1602 55.6% 52.2 s CTS 32
spanish_manual.xml 373605 53373 1 + 1669 55.5% 55.0 s CTS 32
spanish_manual.json 591478 84497 1 + 2642 56.5% 87.0 s CTS 32
image_test 39855 5694 1 + 179 41.4% 5.8 s CTS 32
Table 6
Recall@3 of the gold page for the 54 answerable questions (English questions on each language edition).
Markdown JSON XML
Edition BM25 Semantic Hybrid BM25 Semantic Hybrid BM25 Semantic Hybrid
English 0.89 0.93 0.89 0.91 0.91 0.94 0.80 0.91 0.89
German 0.11 0.87 0.78 0.13 0.91 0.78 0.11 0.89 0.78
French 0.24 0.83 0.80 0.19 0.89 0.83 0.19 0.87 0.74
Dutch 0.20 0.83 0.78 0.19 0.83 0.80 0.17 0.93 0.85
Italian 0.09 0.83 0.80 0.09 0.87 0.78 0.09 0.83 0.78
Spanish 0.15 0.80 0.76 0.17 0.87 0.74 0.13 0.89 0.76
the share of questions, for which the gold page was among
the three chunks (@3) passed to the model.
FortheEnglisheditionofthemanual,thethreemethods
retrieved the gold page for 80 to 94% of the questions. For
the five other editions: BM25 retrieved it for only 9 to 24%
ofthequestions,whereasthesemanticsearchretrieveditfor
80to93%.Intheseeditions,hybridfusionretrievedthegold
page less often than the semantic search alone (74 to 85%),
because chunks ranked only by BM25 displaced semantic
hits in the fused ranking. The mean reciprocal rank of the
firstchunkfromthegoldpagewithinthetopten(MRR@10)
showed the same pattern.
E.2. End-to-end Ablation
Fivecontextconditionswerecomparedwithfourmodels
(Llama3.23B,Qwen2.57B,DeepSeek-R18BandGemma
3 12B) on the Markdown editions of all six languages, with
the question prompt and the judge of the main experiments:

1. no context (empty context), 2) BM25 only, 3) semantic
   only, 4) hybrid (our configuration of the main experiments)
   and 5) oracle (the text of the gold page; for unanswerable
   questions,thethematicallyrelatedpagegiveninthedataset).
   F1 scores were computed with the scoring rule of Sec-
   tion 5.3, with question-level bootstrap confidence intervals;
   paired contrasts were corrected with the Holm method (Ta-
   ble 7).
   Without context, three models declined every question
   andLlama3.23Banswered11%oftheanswerablequestions
   correctly;hybridretrievalthereforeraisedF1by0.65to0.81
   for the English edition and by 0.34 to 0.64 for the other
   editions (all𝑝<0.001after Holm correction).
   For the English edition, the retrieval methods did not
   differ significantly. For the other editions, semantic and
   hybrid retrieval exceeded BM25 by 0.35 to 0.47 (all𝑝 <
   0.001),whereasthehybridandthesemanticretrievaldidnot
   differ significantly.
   The gold page raised F1 above hybrid retrieval by 0.06
   to 0.11 for the other editions, significantly for Gemma 3
   T. Oksanen and J. Gun:Preprint Page 17 of 18

Machine-Readable Operator Manuals over the ISO 11783 Network for Offline, Multilingual LLM Assistance
Table 7
F1 score [95% confidence interval] by context condition (Markdown editions; other languages pooled).
Model No context BM25 Semantic Hybrid Oracle
English edition
Llama 3.2 3B 0.11 [0.03, 0.21] 0.74 [0.64, 0.83] 0.75 [0.67, 0.84] 0.76 [0.66, 0.84] 0.80 [0.72, 0.88]
Qwen 2.5 7B 0.00 [0.00, 0.00] 0.81 [0.72, 0.88] 0.81 [0.73, 0.89] 0.81 [0.73, 0.88] 0.87 [0.81, 0.93]
DeepSeek-R1 8B 0.00 [0.00, 0.00] 0.73 [0.63, 0.82] 0.76 [0.67, 0.85] 0.77 [0.67, 0.85] 0.84 [0.76, 0.91]
Gemma 3 12B 0.00 [0.00, 0.00] 0.78 [0.69, 0.85] 0.82 [0.75, 0.89] 0.78 [0.71, 0.86] 0.84 [0.78, 0.90]
Other five editions (pooled)
Llama 3.2 3B 0.16 [0.06, 0.27] 0.15 [0.07, 0.24] 0.55 [0.44, 0.64] 0.50 [0.39, 0.60] 0.57 [0.46, 0.66]
Qwen 2.5 7B 0.00 [0.00, 0.00] 0.19 [0.10, 0.29] 0.66 [0.57, 0.74] 0.63 [0.54, 0.72] 0.74 [0.66, 0.82]
DeepSeek-R1 8B 0.00 [0.00, 0.00] 0.20 [0.11, 0.29] 0.62 [0.55, 0.70] 0.64 [0.57, 0.71] 0.70 [0.63, 0.76]
Gemma 3 12B 0.00 [0.00, 0.00] 0.22 [0.12, 0.31] 0.67 [0.59, 0.74] 0.64 [0.56, 0.72] 0.73 [0.66, 0.80]
12BandQwen2.57B.Withthegoldpage,thegapbetween
the English edition and the other editions was reduced from
0.13–0.26 (hybrid) to 0.11–0.23 (oracle) so that part of the
gap remained when the retrieval was not a factor.
F. Appendix: Validation of the LLM Judge
The automatic evaluation (judging) relied on GPT-OSS
20B as judge (Section 5.3). To understand the accuracy
of this LLM-as-judge, its verdicts were validated against
humanjudgement,asreportedhere.Theitemsforwhichthe
LLM returned no valid verdict were judged again.
For 147 of the 15552 items of the main experiments
(0.9%),thejudgereturnednovalidverdict:136outputswere
empty,becausethereasoningofthejudgereachedtheoutput
limit of 1024 tokens before an answer was produced, and
11 outputs read “(no)” instead of “no”. These items were
judged again with identical settings, up to five times; 123
items received a valid verdict (105 at the first and 18 at the
secondrepetition).Theremaining24itemswerejudgedwith
the reasoning effort of the judge set to low, which resolved
all of them. The same procedure was applied to the ablation
experiments (Appendix E). No item was therefore excluded
from the analysis.
Four independent non-author raters assessed a stratified
sample of 210 items. The raters had basic understanding of
the machinery type in question, but were not experts on this
particular machine or its operator manual.
The sample comprised six items from each of the 32
combinations of the eight models, the two question types
(answerable,unanswerable)andthetwojudgeverdicts(yes,
no) together with 18 of the 24 items judged with low rea-
soning effort (the other six were answers in the language
of the manual or mixed answers, which were not sampled).
Answersinthelanguageoftheoperatormanualedition(884
items) and answers mixing English with that language (226
items)werenotsampled.Thiswasdonebecausethescoring
rule counts the former as incorrect regardless of the verdict
and the latter contain passages, that the raters could not
be expected to read. For each item, the raters saw a) the
question, b) the expected known answer and c) the model
answer. The model, the language edition, the format or the
verdictofthejudgewerenotshowntotheraters.Theychose
one of three options: match, no match or cannot decide. A
verdict of the judge was compared with the majority of at
least 3/4 raters.
The raters agreed substantially (Fleiss’𝜅= 0.83, 95%
confidence interval 0.77 to 0.88); a three-of-four majority
wasreachedfor201of210items.Thejudgeagreedwiththis
majorityfor188ofthe201items(93.5%;Cohen’s𝜅= 0.87,
95% confidence interval 0.80 to 0.93; Table 8). Weighted to
thecompositionofallevaluatedanswers,theaccuracyofthe
judge was 0.96 (95% confidence interval 0.93–0.98).
Table 8
Verdicts of the judge compared with the majority of four
human raters (201 items with a majority).
Human majority: match Human majority: no match
Judge: yes 102 3
Judge: no 10 86
Most disagreements were rejections of answers that the
ratersaccepted(10of13).Inthestratifiedpartofthesample
(186 items with a majority), for answerable questions, the
judge was wrong in 15.6% of its negative verdicts (7 of 45)
and in 4.3% of its positive verdicts (2 of 46); for unanswer-
able questions, in 2.1% of its negative verdicts (1 of 48) and
in none of its positive verdicts (0 of 47).
Correcting these errors would raise the pooled F1 score
of the English and language-neutral answers from 0.64 to
about 0.67, so the reported F1 scores are conservative. For
theitemsjudgedwithlowreasoningeffort,thejudgeagreed
with the majority for 12 of 15 items. Disagreements among
the raters concerned partially correct or loosely worded
answers (such as incomplete lists, missing specific details
and paraphrased locations).
T. Oksanen and J. Gun:Preprint Page 18 of 18