Citation Freshness for Retrieval-Backed Agents
A retrieval-backed agent should cite a current controlled source or stop. Filename annotations are not enough when the work instruction has a revision stamp.
Direct Answer
Cite a current revision, or refuse to answer from retrieval.
Citation freshness for retrieval-backed agents is the operating rule that a recommendation may use a retrieved passage only when the source is the current controlled revision and the citation can be inspected. If the document is superseded, missing an effective date, or cannot be opened from the citation, the agent should pend the case instead of answering from memory or from a nearby chunk.
That rule is stricter than generic RAG quality. Internal Q&A can tolerate an old explainer. A quality engineer changing a line cannot. The work instruction, deviation, or inspection method either is the released revision or it is not.
Our bias is to treat retrieval as a controlled-document lookup, not as a search box. Vendor citation features help surface a file or a passage. They do not decide whether that file is still the released source.
Floor Practice
Most plants dump PDFs into a store and hope the chunk is current.
The familiar build is fast. Export work instructions, inspection methods, and deviation forms into a vector store. Ask the model to answer from retrieved text. The first demo cites a filename and sounds careful.
On the floor, the failure is a superseded method. A technician asks whether a torque spec changed. The store still holds last quarter's instruction because nobody retired it. The agent cites a PDF that looks official. The traveler gets the old number. Quality finds it after the lot has moved.
A filename is not a revision
WI-4412.pdf can be the retired copy. The released identity is document number plus revision plus effective date.
A similar chunk is not the method
Semantic search will happily retrieve a sister product's instruction or an expired deviation that uses the same nouns.
A citation without a link is theater
If the reviewer cannot open the exact released record, the citation cannot support a process change.
Shop-Floor Packet
A medical-device work-instruction lookup needs five citation fields.
Use a quality-engineering desk at a medical-device plant. The systems of record are the electronic quality-management system, the released work-instruction library, the deviation log, and the traveler or MES record. The agent may retrieve the current method, quote the relevant step, and flag a mismatch against the traveler. It should not release a new revision or close a nonconformance.
The product of the run is a citation packet the quality engineer can check in one screen.
Document identity
Controlled number, revision, title, and the quality-system URL, not only a vector-store file name.
Effective window
Released date, effective date, and whether the document is current, pending, or retired at the time of the run.
Quoted passage
The exact step or table used. Anthropic's citations feature returns the supporting passage from the provided source so a reviewer can see the sentence, not just the file.
Applicability
Product family, line, site, and whether a deviation or temporary change notice overrides the base instruction.
Freshness decision
Use, pend for a newer revision, or refuse because the retrieved source is retired, conflicting, or outside the product boundary.
Implementation
Index released records. Filter before generation. Fail closed on stale hits.
The first integration is the quality system, not the chat UI. If the store cannot carry revision, status, and effective date as filters, the model cannot be asked to invent those fields later.
OpenAI's Responses file-search tool can return file citations and, when requested, the retrieved search results. That is useful evidence. It is still the team's job to keep retired revisions out of the default corpus and to require a status filter before the agent drafts a shop-floor answer.
Index from the system of record
Ingest released instructions and their metadata from the quality system. Do not treat a shared drive export as the source of truth.
Filter before the model writes
Constrain retrieval to current, site-applicable, product-applicable records. OpenAI file search supports metadata filters on the vector store; use them as a control, not as an optional tweak.
Show the passage, not only the file
A reviewer should see the quoted step. Prompt-only 'please cite your sources' instructions are weaker than a citation feature that returns the supporting text.
Pend on conflict
If two current documents disagree, or a deviation overlaps the instruction, route to quality. Do not let the model pick a winner in prose.
Boundaries
Memory, retrieval, and regulated review are three different jobs.
Do not fold this into agent memory. Memory policies at https://solzero.com/blog/agent-memory-policies-for-operational-workflows decide what a run may remember. Citation freshness decides whether a retrieved controlled document is still allowed to support an answer. A remembered exception from last month is not a released work instruction.
Regulated content review at https://solzero.com/blog/content-review-agents-for-regulated-teams is also adjacent, not the same. That workflow reviews outbound language. This workflow checks whether the inbound method is the current one before anyone acts on the floor.
Do not answer from retired hits
A superseded instruction can stay in an archive corpus for investigation. It should not be in the default retrieval set for live questions.
Do not let the model backfill dates
If the source record has no effective date, the packet should say so. Inferred dates become fake confidence.
Do not write the quality system from retrieval
Drafting a traveler comment can wait. Releasing or obsolete-marking a document is a human quality action.
Scoreboard
Measure superseded citations, not retrieval hit rate.
A retrieval agent is working when technicians stop acting on retired methods and quality spends less time reconstructing which revision was used. Useful measures include stale-citation rate, reviewer opens of the cited record without extra search, conflict-pend rate, time from a quality-system revision to corpus update, and shop-floor corrections tied to a wrong instruction.
Do not celebrate more documents in the store. An index that still serves retired revisions is a faster way to quote the wrong method.
The SolZero take is that citation freshness is a document-control problem that happens to use retrieval. If you want a workflow review for a plant that already has a vector store and still argues about which revision was current, start at https://solzero.com/#how-it-works.
Further reading