An AI tool ignores part of a long document: check context and retrieval separately

Find the source passage yourself

Open the approved document edition and locate the relevant wording. Record the title, version, page or section and surrounding qualification. Check whether a later appendix, table note or diagram changes the main rule.

Do not use a remembered answer as the reference. If the document is ambiguous or contradictory, resolve that editorial issue first. An assistant cannot reliably turn conflicting policies into one authorized answer without knowing which applies.

The fact-checking workflow explains source ownership and claim review. Here the goal is a narrower trace: what happened to one passage between the document and the answer?

Distinguish capacity from selection

Google’s long-context guidance describes bounded context and model-specific capacity arrangements. A large supported window is capacity, not evidence that the application included the entire uploaded file in a particular request.

Google’s File Search documentation describes importing, chunking and indexing material, then retrieving relevant information as context. In that design, the answer may use selected passages rather than reading every page for every question.

Ask the maintainer which path your tool uses: direct document input, extracted full text, retrieval, summarization or a combination. Avoid assuming that a file-size allowance is the same quantity as model context capacity or per-answer retrieved content.

Inspect extraction before tuning retrieval

Check tables, scans, footnotes and multi-column layouts in the extracted representation. A document can look readable to a person while producing jumbled text. An image-only instruction may need an appropriate approved transcription or supported visual-processing route.

Do not remove access controls to make an internal manual easier to ingest. The knowledge-boundary guide explains why a useful source still needs the right audience.

For a system where the page or file is absent altogether, use the ingestion diagnostic for the earlier discovery and processing stages. This guide begins when the source is present but the answer misses meaningful content.

Review chunk relationships and selection

A passage’s meaning may depend on its heading, preceding definition or later exception. Check whether those relationships survived chunking. A sentence containing this condition is difficult to interpret if the condition’s definition was split away.

Use the system’s available retrieval diagnostics to compare selected passages with the expected ones. Record filters, source edition and any ranking configuration. Do not simply increase the number of retrieved chunks until the answer changes; unrelated material can add competing versions and make the cause harder to identify.

For the fictional manual, a useful small experiment would compare the original question with one that explicitly names the harsh-condition section. If that retrieves the exception, it supports a selection investigation. It does not establish that every visitor should know the document’s internal terminology.

Confirm the actual request boundary

Token-counting documentation distinguishes preflight request size from reported response usage. Ask for the supported counts and a minimized trace where available. Include instructions and conversation history, not just the document text.

If the request is truncated, identify which layer did it and what policy it used. A frontend upload limit, extraction cap, retrieval limit and model context limit require different corrections. Keep credentials and unrelated personal content out of the diagnostic record.

Where the hosted product exposes no request trace, state that limitation. You may be able to compare cited passages and controlled synthetic cases, but you cannot honestly certify that the whole file was supplied.

Test interpretation with the relevant passages

In a safe evaluation environment, provide the approved main rule and exception together and ask the same question. Define the expected answer before running it. If the response still ignores the qualification, the generation instructions or model behavior deserve review.

A successful isolated answer is useful diagnostic evidence, not acceptance of the entire system. Return to the actual retrieval path and verify that it supplies both passages for ordinary user wording. Include a nearby question where the exception does not apply, so a fix does not over-apply it.

Keep section terminology in the test record, but also retain the ordinary wording a reader would use. A system that finds the exception only when given its exact heading may still fail its real audience. That gap is a useful retrieval acceptance case, rather than a reason to require every reader to study the manual first.

Choose the smallest justified correction

Possible corrections include clearer source headings, repaired extraction, a better edition filter, preserved neighboring context or an adjusted request budget. Select the one supported by the trace. A model upgrade may be appropriate only after capacity or interpretation has actually been implicated.

Retain the question, expected passages, actual selected context and result as a repeatable case. Your next step is a passage-level evidence chain. It turns the vague complaint that AI ignored the file into a specific stage that can be repaired and verified.