Prepare documents for RAG

A knowledge search returns passages from your documents. Prepare each passage so your persona can understand it without the surrounding pages: name the subject, keep conditions beside the facts, and preserve the source. Start with the document template, or jump to the task you need. All examples on this page are fictional. Choosing which material a persona should access? See Set the knowledge scope.

​

Organize your documents

Scenario: You have several manuals, policies and FAQs to upload. Aim: Each file covers a useful topic and keeps its conditions together. Example Instead of one “Customer handbook”, create:

• returns-policy.md: deadline, eligibility, exclusions and process.
• installation-guide.md: requirements, steps and completion checks.
• billing-faq.md: payment methods, invoices and billing questions.

Separate regional policies when their rules differ substantially. Use clearly labeled sections when the differences are small enough to explain together. Quick check: Can someone understand the topic and its important restrictions from this file without hunting through unrelated documents? Anam processes uploaded documents into searchable passages. Inspect a retrieved passage: even a well-organized file can be split at an unhelpful point. There is no universal file count or document length that guarantees accuracy.

​

Write clear, complete facts

Scenario: The persona finds a passage but misinterprets what it applies to. Aim: A fact remains clear when read outside the full document. Example Don’t: “Support resolves issues within two hours.” Do: “The Business plan has a two-hour initial response target during support hours. This is not a resolution deadline.” The second wording preserves what the source actually measures. Similarly, listing supported configurations does not establish which configuration suits every user. Include decision criteria only when the source provides them. Quick check: Does the passage identify the relevant plan, version, audience or region? Could a reader mistake a target for a guarantee? Short passages can lose their original context. Clear names and nearby qualifications help them remain interpretable when retrieved alone.

​

Keep critical rules with the facts

Scenario: The persona retrieves a deadline, price, or procedure without the restrictions that govern it. Aim: Each passage includes the conditions needed to use it correctly, even when retrieved on its own. Example: a fictional returns policy Weak passage

Returns are accepted within 30 days.

Self-contained passage

For online purchases in Region A, unused items can be returned within 30 days of delivery. Personalized items are excluded.

Include those conditions wherever the document explains the deadline or return procedure. Preserve this useful repetition when removing duplication.

​

Check how context reaches the model

Inspect the passages returned by search to check that rules and exceptions stay together. A passage is only as complete as the text around it in your source file. Where a restriction matters, put it in the same paragraph, table row, or step as the claim it qualifies, rather than relying on a nearby section to travel with it. Carry the identifying details a reader would need — product, version, region, effective date — into the text itself, since a filename or a document property is not part of the passage the model receives. Check the result rather than assuming: retrieve the passage and read what actually comes back.

​

Enforce behavioral guardrails in your application

“Personalized items are excluded” belongs beside the returns policy. A behavioral rule such as “Never issue a refund without authorization” also needs enforcement in the application. Check authorization in the code that performs the refund; repeating the rule in documents or prompts does not guarantee compliance. Quick check: Retrieve the pricing, eligibility, and procedure sections separately. Does each returned passage include the conditions needed to use it correctly? Include this check in your knowledge base tests.

​

Remove duplication

Scenario: A manual, FAQ and sales sheet repeat similar information. Aim: Remove repeated wording while preserving unique evidence. Example An overview says: “Business customers receive priority support.” A policy adds: “The initial response target is two hours during support hours. Training requests are excluded.” Don’t: Replace both with the overview sentence. Do: Retain the response target, hours and exclusion in the main support policy, with its source reference. Quick check: Compare the old and new versions for every number, limit, exception and procedural step. Record why any unique information was removed. Preserve the original files separately. Keep short, repeated restrictions that make individual passages complete. See Keep critical rules with the facts for the distinction between necessary context and redundant copies. If a tool response still repeats overlapping passages after source cleanup, contact support with the query and repeated passages. Editing documents does not fix every form of duplication. Reusing a document in separate searches is not itself a defect.

​

Preserve formatting and steps

Scenario: You are uploading a PDF, price table or step-by-step procedure. Aim: The extracted text preserves the relationships shown in the original. Example Don’t: “Install the update and restart.” Do: “Before updating, create the backup required by the installation guide. Install the update, restart, then run the documented verification check.” Use the second wording only when those steps are supported by the source. Do not invent missing steps to make a procedure seem complete. Quick check: Look for missing columns, misplaced footnotes, broken reading order and images containing essential text. Use a supported format that preserves these details; clean Markdown or text may help, but still check the result. Flag unreadable material for correction instead of guessing its contents. For an update question, test whether the persona includes the prerequisite as well as the main action.

​

Handle conflicting information

Scenario: Two sources give different answers to the same question. Aim: Preserve valid differences and make unresolved conflicts explicit. Example Source A gives a 14-day return period. Source B gives 30 days. Don’t: Publish “Returns are accepted within 30 days” for everyone. Do: If the sources cover different regions, label each rule with its region. If both cover the same region and effective period, record the conflict and withhold a definitive deadline until the policy owner resolves it. Keep internal discussions and review notes outside the persona’s knowledge folders. In the reference document, explain any remaining uncertainty beside the affected fact. For example: “The return period for UK purchases is awaiting confirmation because the two source policies give different deadlines.” Quick check: Have you identified the conflicting rules, their sources, who they apply to, and who will resolve the disagreement? Record policy effective dates separately from the dates you reviewed the documents.

​

Next steps