Designing a Knowledge Base and RAG for WhatsApp AI

Inventory policies, FAQs, catalogs, procedures, product guides, and service information, then assign an owner and review cycle. Mark drafts, approved versions, validity periods, markets, languages, and audiences. Retrieval cannot repair contradictory source material reliably.

Keep dynamic customer facts outside documents. Order status, stock, price, appointment capacity, and balances belong behind live tools. The knowledge base explains stable rules and guidance; systems of record answer changing transactional questions.

Extract clean text while preserving headings, tables, lists, page references, source identity, and version. Reject unreadable or unsupported files visibly. Store a checksum so repeated uploads do not create duplicate knowledge and changed content can trigger controlled re-indexing.

Chunk around semantic boundaries rather than arbitrary character counts. Include enough local context to explain a rule while avoiding unrelated sections. Carry tenant, document, language, status, and permission metadata into every chunk so later filtering remains enforceable.

Filter by verified tenant and access policy before ranking candidates. Combine semantic retrieval with keyword or metadata matching where product codes and exact terms matter. Use query rewriting carefully, because an expanded query must not change customer intent.

Return a focused evidence set with source metadata and relevance signals. When evidence conflicts, is stale, or lacks the requested fact, route to clarification or a human. Increasing the number of chunks often adds distraction rather than certainty.

Tell the model which statements must be supported and how to handle missing information. Separate retrieved passages from instructions so documents cannot override system policy. Ask for concise customer language and retain source references internally for review.

Validate structured claims when possible and block answers that require evidence but received none. For sensitive workflows, expose the supporting source to teammates or customers in an appropriate form. Traceability turns a mysterious answer into an improvable operational decision.

Build questions from real conversations and label the correct source, expected facts, and acceptable abstention. Measure retrieval recall, ranking quality, factual support, tenant isolation, and answer usefulness separately. A fluent answer can hide poor retrieval.

Re-run evaluation when documents, embedding models, chunking, prompts, or filters change. Review unanswered questions as a content backlog. A healthy knowledge system improves through publishing discipline, observable misses, and controlled revisions rather than uncontrolled prompt growth.