You upload several reports, add a technical manual and ask an AI assistant for a complete analysis. The answer is clear and well structured. But has the assistant considered every document, or mainly the passages it selected for your question?
A successful upload confirms that a file was accepted. It does not, by itself, establish that every page contributed to the answer. To assess coverage, you need to distinguish the files available to the application from the information actually used for a particular response.
Available files and active context are different things
The context is the information supplied to a language model when it generates a response. Depending on the application, it can include instructions, conversation messages, document text and results returned by tools. The context window limits how much information the model can handle in that operation.
An application can make a larger collection available through search or file-reading tools. The assistant may retrieve relevant passages, open additional sections or process documents in stages. The precise workflow depends on the product and task.
Persistent memory is another mechanism. Where supported, it can preserve selected information between interactions; it should not be treated as a complete copy of every uploaded file. Anthropic’s context-engineering guidance describes approaches such as retrieval, summaries and external notes for managing information across tasks.
| Concept | What it means for document work |
|---|---|
| Uploaded files | Material accepted by the application and potentially available for processing. |
| Active context | Information supplied to the model for a particular response. |
| Retrieval | A search process that selects material from a larger collection. |
| Persistent memory | Selected information retained across interactions, where the product supports it. |
These distinctions help explain why a file can remain available in a project without its entire contents appearing in every response.
How document retrieval works
Retrieval-augmented generation, usually called RAG, combines a search step with answer generation. A common implementation divides documents into smaller passages, searches for relevant material and supplies selected results to the model.
Anthropic documents this approach for Claude projects: when project knowledge approaches the context-window limit, RAG can activate, allowing Claude to search uploaded material rather than load all project content at once. This is a documented Claude workflow, not a universal description of every AI application.
Search is useful when you need a specific fact in a large collection. However, finding evidence for a targeted question and reviewing an entire collection are different tasks. Our practical conclusion is that an exhaustive request needs a way to check coverage, rather than relying only on the relevance of retrieved passages.
A complete-looking answer can still have gaps
Consider a fictional example: you upload three supplier proposals and ask which offers the best value. An assistant compares headline prices and recommends the cheapest option. Yet one proposal includes installation, another bills it separately, and the third has a maintenance charge in an appendix.
A useful comparison must reconcile those conditions. If the output covers only the prices, its polished presentation does not resolve the missing information.
For this task, define the fields before asking for a recommendation: equipment, installation, recurring charges, exclusions, warranty and delivery schedule. Ask for evidence from each proposal and have missing fields marked explicitly.
“Not found in the reviewed material” is a more cautious finding than “not included in the proposal.” Establishing absence generally requires broader coverage than finding one relevant passage.
A larger context window does not guarantee complete understanding
A model’s input capacity and its ability to use every relevant detail are separate questions. The 2023 study Lost in the Middle found that the tested models’ performance on retrieval and multi-document questions could change substantially with the position of relevant information, including poorer results when it appeared in the middle.
That study concerns particular models and tasks from its period. It does not establish the performance of every current system. It does illustrate why an advertised context size alone is insufficient evidence of reliable document analysis.
For a broader approach to evaluating models, see our guide to choosing an AI model beyond benchmark rankings. In document work, test the information you need extracted and the time required to check the result.
A practical method for checking document coverage
The following workflow is an editorial recommendation for making document analysis easier to review. It cannot guarantee that every relevant detail will be found.
1. Define the document set
Keep your own list of filenames, versions and dates. Use descriptive names and identify which version should take precedence. Ask the assistant to report which named files it can access; compare that report with your list.
A file inventory is a starting point. Listing a filename does not prove that the file’s contents were inspected.
2. Specify the evidence you need
Replace “analyse everything” with a defined task. For a comparison, specify the same fields for each document. For a report review, identify the sections and questions that matter.
This gives you a concrete basis for detecting omissions.
3. Request a result for each document
Before requesting an overall synthesis, ask for findings by file, with page or section references where available. Separate missing information from inaccessible or unreadable material.
If the tool cannot process a document completely, ask it to state that limitation. Check a few references yourself, especially those behind the main conclusion.
4. Compare the findings and revisit the originals
Use the document-level results to build the synthesis. Then return to the original passages behind contradictions, unusual figures and decisive conditions.
Summaries are helpful working material, but they can omit details. Keep the originals available during the comparison.
5. Make unresolved gaps visible
Ask the final output to identify documents or sections that could not be checked, uncertain references and questions left open. If the application exposes search results or reading activity, use those records as additional evidence of scope.
The assistant’s own account of what it reviewed is useful, but it remains a claim to check.
A prompt for reviewing several files
Analyse the following named documents: [filenames].
First, report which files you can access and any processing limitations.
For each document, extract [specified fields] with supporting page or section references where available.
Distinguish information found, information not found in the material reviewed, and material you could not inspect.
Compare the findings only after producing the document-level results. Recheck the original passages behind important differences.
Do not claim complete coverage unless you can substantiate it, and do not invent references.
This prompt makes the scope explicit. It cannot force a product to expose files, text or processing capabilities it does not provide.
Match the review to the task
For a narrow question, targeted retrieval may be an efficient starting point. For a comparison across files, document-level extraction makes omissions easier to detect. For a comprehensive review, request section-level coverage and verify the passages that determine the outcome.
Before document findings lead to an external action, consider the authority you have delegated. Our guide to AI agent permissions for emails, files and purchases explores that related decision.
The useful question is whether the answer shows enough evidence and coverage for your task. A filename, a fluent summary or a large context window cannot settle that question on its own. A reviewable result identifies the material used, supports its conclusions and makes gaps visible.
Sources
- Claude Help Center — Retrieval augmented generation (RAG) for projects — Product-specific explanation of search over project documents.
- Anthropic — Introducing Contextual Retrieval — Explanation of passage retrieval and the importance of surrounding context.
- Anthropic — Effective context engineering for AI agents — Context management, retrieval and persistence strategies.
- Liu et al. — Lost in the Middle: How Language Models Use Long Contexts — Research on how the tested models used information in long inputs.