The factory's memory
Generative AI for turning scattered documentation into operational knowledge
IN REAL SETTINGS: INDUSTRIAL APPLICATIONSRESEARCH
Beatriz, GPT Astra 6, Claude Opus 5.5
11/18/20269 min read


Picture a seemingly simple question: which ingredient specification should be used for an order? The folder contains two specification sheets, a later email and an exception accepted for one particular customer.
An agent could summarise all four documents beautifully and still recommend the wrong combination: the old sheet, another customer's requirement or an exception that no longer applies. The problem starts long before the model writes anything: how has the context needed to interpret those documents been preserved? It is one thing to have the information and quite another to know how to use it. In businesses large and small, the right answer may depend on the product, the supplier, the date, the processing conditions and who approved a change. Much of that judgement is tacit knowledge. Stripping away those relationships to produce a cleaner explanation removes precisely what makes the data useful. This is where generative AI deserves a closer look: in the work of retrieving, structuring and cross-checking knowledge that someone will later have to use.
What I mean by operational memory
I use the factory's memory as a working definition: the body of sources, criteria and documented decisions that makes it possible to reconstruct why something was done in a particular way and to judge whether that criterion still applies. It is not simply a matter of finding a procedure. I need to be able to tell which version is current, locate a correction and recognise the limits of an authorisation. A folder stores files; supporting decisions means preserving the relationships between them.
We should not ask too much of the idea. Some professional knowledge lives in experience and in practices that nobody has fully described. Uploading documents to a system does not capture that know-how, nor does it turn every habit into a valid procedure. What it can do is show where a task relies on scattered information, which contradictions someone has to resolve and which criteria need to be documented more clearly. That organisational work is part of the project; it is not an irritating preliminary to be got out of the way before artificial intelligence can begin.
Retrieval-augmented generation (RAG) offers a technical basis for combining answer generation with information retrieved from an external repository. The work of Lewis and colleagues helps to explain this architecture and to distinguish it from the knowledge stored in a model's parameters. Whether it helps with a particular industrial query is a separate question, and one that has to be tested case by case. [1]
The first design questions are distinctly unglamorous: which sources may be consulted, what it means for them to be current and how the system should respond when they disagree. Nor does choosing between running the system locally and using external services answer them. The place where a file is processed and the quality of its contents are related, but separate, decisions.
A recent case in food engineering
AgentXtract offers a concrete example. The study by Zhou and colleagues organises the extraction of the dielectric properties of foods, which govern their response to microwave and radio-frequency heating, and uses the resulting database in a query prototype with document retrieval. People still have clearly defined tasks: selecting sources and reviewing ambiguous records. The evaluation distinguishes between locating a record and extracting its values correctly; without that distinction, it is difficult to interpret what it means for the system to work. [2]
According to the published results, the database contains 4,276 records for 142 food materials, extracted from 64 articles. The team reports roughly 30 minutes of automated extraction and 9 US dollars in calls to the model selected to build it. [2, §3.2]
Penn State's announcement of 16 October explains how the work was divided: experts selected and assessed the sources, while agents processed the scattered information. According to the same announcement, the research received start-up funding from Penn State and from Anthropic's AI for Science programme; Anthropic was one of the providers whose models were compared. The article is available online ahead of print and has been assigned to the March 2027 volume of Journal of Food Engineering. The announcement and the article are not independent pieces of evidence; they belong to the same chain of evidence. [3]
The published data schema allows for conditions such as temperature, frequency and moisture content, and records the provenance of each entry. It also distinguishes values read directly from the sources from those derived from equations. This is a significant design decision: keeping a number means keeping what it represents and how it was obtained. The presence of those fields does not, of course, mean that every record is complete. [4a]
The case suggests a transfer hypothesis: organised extraction could make technical documentation easier to reuse, provided that the required data, their relationships and their checks are defined in advance. That hypothesis deserves its own trial in each workflow. Extracting properties from scientific articles does not show that a system can resolve discrepancies between suppliers, approve a specification sheet or interpret a shop-floor incident.
These figures require some discipline when discussing costs. Model call charges alone do not account for corpus preparation, development, specialist review or maintenance. Nor does the automated extraction time cover that work. Access to the code makes it possible to examine the design; reproducing the results also requires the corresponding inputs and evaluation conditions. [4b]
A note on the data. The figures above are those of the published article, which attributes them to its Supplementary Table S1. The file distributed as S1 describes itself as an updated version and contains different counts: 4,451 records, 145 materials and 52 sources. The documentation consulted does not explain the difference; I flag it for transparency, and this analysis is not an independent reproduction of the results. The supplements are available from the article page.
Keeping the data without erasing its history
I would carry a three-layer separation over into industrial design. First comes the source that the organisation recognises and holds. Next is the extraction or interpretation produced by the system, which must be correctable without altering that source. Finally, there is the decision to use it: who accepts the result, for what purpose and under what restrictions. The three layers are related, but each must retain its own identity so that a correction does not retrospectively rewrite what happened.
Suppose a document gives a concentration. We need to know the unit, the product to which it refers and the conditions under which it was measured. If it contains an instruction, its version and scope matter. If it records an exceptional decision, the reason for the exception and its limits need to be preserved. Blank fields, contradictions and values awaiting confirmation need to be dealt with explicitly. Filling them in merely to complete a table would make things look tidy at the expense of knowledge.
Some of this control can be automated. The AgentXtract code I examined contains rules for flagging certain inconsistencies. [4c] Their relevance outside that domain should not be assumed: each workflow needs its own rules and competent review. A match between identifiers or a well-defined unit conversion can be checked without asking the model to improvise an explanation.
To justify an answer, I would retain the source passage and a record of the transformation applied. A citation may lead to the right file and still accompany the wrong interpretation. The generative AI profile published by the US National Institute of Standards and Technology (NIST) recommends checking sources and citations, documenting human contributions and avoiding extrapolation from narrow evaluations. It is a voluntary methodological reference, not an accreditation of the system or a food safety standard. [5]
Two applications that deserve a bounded trial
The first would be to compare specification sheets with supplier documentation. In this hypothetical scenario, the system would draw up a table containing the fields that matter to quality or purchasing, state where each value came from and flag differences between versions. The acceptance criterion would need to distinguish a difference in wording from a change that affects the intended use. Initially, the output would be no more than a proposal for review; approval and master-data updates would remain matters for the responsible professionals.
Here, finding a conflict may be more valuable than producing a neat summary. If one sheet gives a value and another contradicts it, the useful job is to bring the discrepancy into the open, establish whether both refer to the same material and route the issue for clarification. Simply choosing the value that appears most often, or the document whose filename looks most recent, would be shortcuts without sufficient justification.
The second application would be retrieving past incidents. Faced with a hypothetical query about a packaging defect, the system would look for documented episodes and allow the product, equipment, observed conditions, actions taken and recorded outcome to be compared. Similar wording can help to identify candidates; we then need to establish whether the relevant conditions genuinely match. An action that was proposed, one that was carried out and one whose effectiveness was verified must remain distinct facts.
These are proposed applications, not implemented cases or measured benefits. Their commercial value would depend on how often the task occurs, the quality of the sources and how much work they genuinely take out of the process. If there is little documentation and it is well organised, ordinary search or a structured query may be enough. Any additional complexity needs a justification that stands up to that comparison.
Measuring whether the work actually gets done
To assess a pilot, I would start with one specific document-based decision and a set of queries whose answers a competent reviewer could check. I would compare the current way of working with the proposed approach using the same inputs and acceptance criteria. The evaluation should also include contradictory documents, superseded versions and questions that cannot be answered from the material available. Admitting that the evidence is insufficient is part of a useful answer.
My minimum evaluation framework would look like this; it is a working proposal, not a validated standard. First, does the system retrieve what is relevant? On known queries, I would check that the document, version and context are right; that would show whether retrieval is fit for the task. Second, does the answer hold up? Each claim needs to be checked against the cited passage, with errors classified by type and severity, before deciding which uses are acceptable. Third, does it recognise its limits? Missing information, contradictions and out-of-scope questions show when it should ask for clarification rather than fill in the gaps.
Two further questions are just as practical. Fourth, does the system reduce the total workload rather than simply the search time? Preparation, review and correction through to acceptance all need to be counted before claiming an operational improvement. Fifth, can it be maintained? Update costs, clear ownership and a change log show whether the capability can outlast the initial enthusiasm.
Generation time would be only one part of the measurement. I would also count the work that the tool shifts onto the reviewer and the queries eventually resolved by other means. To estimate the cost per query successfully resolved and accepted, failed attempts and an explicit allocation of preparation and maintenance would have to be included over a defined period. Otherwise, one could make the project look efficient simply by leaving out the difficult parts.
Nor would I set a universal accuracy target. An incomplete answer to a secondary query and a mix-up between specifications have very different consequences; the evaluation must allow them to be seen separately. Acceptance criteria should be agreed by whoever understands the intended use and is accountable for it. A reassuring average can hide precisely the errors that matter.
Expert knowledge needs maintenance too
In the design I propose, people have defined work: selecting sources, resolving conflicts, accepting exceptions and retiring criteria that are no longer useful. The cost and responsibility involved must be acknowledged. Leaving someone at the end of a chain of answers, with neither the time nor the information to review them, would turn oversight into rubber-stamping.
I would therefore start with a small workflow whose document dependencies are understood. A family of specification sheets, a recurring technical query or a set of incidents offers a way to learn which context needs to be kept and which checks are worth automating. Scaling up would follow the evidence, rather than serve to make the project look ambitious. Finding that sources are missing, that a criterion is disputed or that a simpler approach would do the job would also be a useful outcome.
The factory's memory matters because it makes the criteria we work with open to question and review. Its value lies in enabling someone else to find the source, understand the decision and correct it when conditions change. What has been preserved in this way can support an AI answer. Eloquence cannot recover what was lost along the way.
Sources and scope of the reading
Documentary review undertaken on 20 October 2026 and checked again on 18 November, including a reading of the full article and both official supplements. This analysis uses a research case as an illustration and considers its possible transfer; it is neither a systematic review nor a replication. Figures are attributed to the published article, with the note above on the supplement.
Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020. Official record. Used as architectural grounding.
Zhou, X.; Zhang, Y.; Sablani, S. S.; Wang, S.; Tang, J. Agentic AI for food property data: From literature to database to new discovery. Journal of Food Engineering, 424, 113327, March 2027 volume; available online 12 September 2026. DOI, full text hosted by the authors, sources supplement (.docx download) and supplementary database (.xlsx download). Read section by section, with the database counts checked; discrepancies remain to be clarified.
Penn State. Can AI help discover new food physics? Researchers pair AI with human expertise. Institutional web page, published 16 October 2026. Announcement.
zhouxu-lab. agentXtract. Public code, version e733e7c, 17 September 2026. Repository. Targeted static review: 4a, data schema; 4b, data availability; 4c, programmed checks. The system has not been run.
National Institute of Standards and Technology (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1. Official document. Targeted reading of section 2.2 and actions MS-2.3-002 and MS-2.5-001/002/003/005.
Use of AI. The documentary research was carried out with Astra xhigh (OpenAI). The writing is the author's. Proofreading, accuracy checks and the English translation were done with Claude (Anthropic).
Research by: Beatriz Vallina, PhD
Thesis Supervisors: Roberto Cervelló, Prof.PhD & Juan José Lull, PhD
Institution: Doctorate in Agrifood Economics, Universitat Politècnica de València
© 2026. All rights reserved.
