Internal documents for AI: choose full text or search by document
Choose full text or search for each document by update frequency and the scope needed for an answer.
Key points
- Split policy and repair-record inputs
- Record updates and the scope needed
- Define when to switch and return
Overview diagram in preparation
Contents
Use different methods for policies and repair records
A&A perspective
There are two ways to supply internal documents to the AI that writes an answer: provide the full text, or search first and provide the relevant parts. Keep full text as an option for short policies; consider search for repair records needed only in narrow slices. Both must satisfy the AI's capacity and your confidentiality requirements.
A&A perspective
Suppose the manuals you send together on every request have grown, and answers now take longer. Moving every document to search introduces another concern: whether clauses defining exceptions will be found. Separating methods by document characteristics can narrow the work you need to rebuild.
From the sources
Brand.ai, which provides software for applying brand strategy in day-to-day work, says in an Anthropic customer story that it values Claude's ability to process entire brand guidelines at once.
From the sources
The company also uses two retrieval systems: one surfaces relevant brand insights, and the other maintains context in conversations. Its platform combines the ability to handle whole documents with mechanisms for finding relevant information.
A&A perspective
Automate & Augment (A&A), the publisher of this article, applies that combination to supplying internal documents. We borrow the document-reading mechanism. This does not mean a small company can reproduce brand.ai's product scale or results.
Hypothetical example
Consider a fictional equipment-maintenance company. It supplies the body and exceptions of its travel-approval policy together. From past repair records, it supplies entries found by equipment model and symptoms. The policy and the history do not need the same input method.
A&A perspective
For confidential material, first establish where the original documents and search data may be stored. An index is data organized for searching. If the index cannot be hosted outside the company, choose a method within that constraint. Separately check whether retrieved excerpts may be sent to the AI.
A&A perspective
The related article “How much customer data to give an AI contractor: decide 3 tiers first” covers which information may be supplied. This article decides which portions to retrieve from documents already cleared for use.
Record update frequency and the scope each answer needs
A&A perspective
For each document, record how often it changes and how much is needed to answer typical questions. Using these two factors is A&A's proposal. The sources do not provide numerical thresholds for this decision.
From the sources
Anthropic's tool-design article gives the example of finding one person in an address book. It argues that returning every contact for the agent to inspect consumes limited context space with irrelevant information.
A&A perspective
That observation makes it worth examining not only whether the material fits, but how much of it an answer needs. A reviewer compares typical questions with the original passages needed to answer them. A list of citations generated by the AI is not proof that it considered every exception.
Hypothetical example
In the maintenance-company example, the travel policy's entry reads: updated when revised; the whole document, including exceptions, is needed. The repair-history entry reads: added to after each job; only similar symptoms on the same model are needed. These are fictional workflow examples, not research findings.
A&A perspective
For frequently changing documents, also compare the work needed to supply the latest version. Retrieving the full current document each time requires work, as does propagating changes into a search index. Adding search alone does not establish freshness; decide who checks that updates have been reflected.
Hypothetical example
Even a short work schedule that changes daily can remain a full-text candidate if the answer needs the whole schedule. Conversely, a thick manual that rarely changes can be a search candidate when only the pages for a particular equipment model are needed.
A&A perspective
Using update frequency alone can assign these examples the wrong way around. Volume still matters. If the full text exceeds the AI's input limit, consider splitting it by business task or using search.
A&A perspective
Record the document name, update frequency, scope needed for an answer and current input method. If it is unclear whether the whole document or only part is needed, keep that uncertainty explicit. Before commissioning search for those documents, a staff member gathers actual questions and the pages needed to answer them.
| Document | Updates and required scope | Method and review condition |
|---|---|---|
| Travel-approval policy | Updated when revised; body and exceptions needed together | Keep full text within capacity; a person checks necessary clauses |
| Repair history | Added to after each job; matching model and symptoms needed | Try search; return if a required record is missed, subject to capacity and version checks |
| Short work schedule | Changes daily; the whole schedule is needed | Keep the latest full text as a candidate; frequency alone does not decide |
Define when to switch to search and when to return
A&A perspective
On the same row of the document allocation sheet, record conditions for moving to search and returning to full text. Consider moving documents whose useful portions are narrow and whose full-text delivery has become burdensome. Before switching, check that the needed passages can be retrieved from the latest version.
From the sources
Anthropic recommends beginning tool design with a few tools aimed at high-impact workflows. It describes evaluating accuracy alongside execution time, tokens—the units related to how much text the AI processes—and errors, among other measures.
A&A perspective
A&A proposes trying the same business questions with both full text and retrieved excerpts. A reviewer first identifies the required text and exceptions, then checks which document version the answer used. Compare waiting time, actual usage charges, and the human work of checking and updating. Benefits remain unmeasured until the trial.
Hypothetical example
For repair history, a switch condition could be: records for the target model can be retrieved, including the latest additions. A return condition could be: a required historical record does not appear in search. Failure to retrieve a record is not evidence that it does not exist.
A&A perspective
A short policy whose applicability conditions and exceptions must stay together is a candidate for continued full-text input, within capacity limits. The AI can still overlook material even when given the whole text. A staff member checks the original when the necessary clauses cannot be confirmed in an answer.
A&A perspective
Returning to full text also requires a usable current version. If the full document does not fit, or its latest version cannot be confirmed, return to human checking of the original. Do not keep automated answers running when neither method can be verified.
Hypothetical example
An outsourcing brief could cover search over repair history and checking that new entries are reflected. Decisions about revising the policy stay inside the company. Naming the reviewer in the sheet makes it possible to agree separate roles for building the mechanism and deciding whether the source material is correct.
A&A perspective
If the categories of information themselves need organizing first, start with the related article “Stop repeating yourself to AI: split company info into 3 parts”. The sheet in this article follows that work: it assigns an input method to each document and keeps the scope of a change narrow.
Review full text and search choices document by document
A&A perspective
Supply policies and repair records differently according to the scope of information each answer needs.
A&A perspective
Recording update frequency and required scope helps compare replacing full-text inputs with maintaining search data.
A&A perspective
Keeping switch conditions, return conditions and the reviewer on one sheet makes it possible to decide changes document by document.
Supply policies and repair records according to the information each answer needs. Compare the work of replacing full text with updating search data. Record when to switch, when to return and who checks the source, so each document can be changed separately.
FAQ
Will moving to search always reduce usage charges?
No. Beyond the amount sent to the AI, search execution and index updates have costs and require work. A&A proposes comparing actual charges and human work using the same questions.
Does supplying a full policy prevent missed exceptions?
Not necessarily. It removes the retrieval step that could omit an exception, but the AI may still overlook it. For important answers, a staff member checks applicability and exceptions in the original.
What if a document's update frequency is unknown?
Keep it marked as unknown rather than guessing. Check the most recent revision and who maintains the latest version. Even when switching to search, you still need to check that the latest version is reflected.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Brand.ai uses AI to make brands more human with Claude
Anthropic · Publication date not stated
Accessed 2026-10-30 - Writing effective tools for agents — with agents
Anthropic · 2025-09-11
Accessed 2026-10-30