← Voltar aos guias

Build a Team Knowledge Assistant That People Trust

Start With Decisions, Not Documents

A useful internal knowledge assistant does more than summarize a folder of PDFs. It helps employees answer recurring operational questions: which deployment procedure applies, who owns a service, what the current security exception process is, or how to configure a customer integration. Begin by collecting 30 to 50 real questions from support tickets, onboarding chats, incident reviews, and internal search logs. Group them by intent, then define the expected answer format for each group.

For example, a production-change question may require a short answer, the current policy version, source links, and an escalation owner. A technical troubleshooting question may require ordered steps, prerequisites, and a warning when evidence is incomplete. This schema becomes the assistant's contract. It also prevents a common failure mode: building an impressive demo that produces vague answers no team can safely act on.

Design Retrieval Around Trustworthy Chunks

Retrieval quality usually matters more than selecting the largest model. Normalize your source material before indexing it. Convert documents into clean text, preserve headings, remove repeated navigation, and attach metadata such as source URL, owner, team, publication date, document type, access level, and last-reviewed date.

• Chunk by semantic section, not arbitrary character counts. A runbook section, policy clause, or API endpoint reference should stay intact where possible.
• Keep chunks roughly 300 to 700 tokens, with a small overlap only when a procedure spans a section boundary.
• Store the heading path with every chunk, such as “Engineering > Releases > Rollback procedure.”
• Include version and freshness metadata so a superseded policy cannot silently outrank its replacement.
• Index structured sources separately. A service catalog, ownership directory, or incident database benefits from field-aware retrieval instead of plain text search.

Use hybrid retrieval: lexical search for exact identifiers, acronyms, ticket IDs, and error messages; vector search for semantic matches; and a reranker to select the best final evidence. Retrieve more candidates initially, then rerank the top 20 to 50 results. This is particularly valuable when internal language includes product names and abbreviations that embeddings alone may mishandle.

Make Citations a Required Output

In the generation prompt, instruct the model to answer only from provided evidence and cite each material claim using source title, URL, and section. Require it to say that it cannot verify an answer when the retrieved context is insufficient. Do not treat this as a cosmetic UI feature: citations let employees validate recommendations quickly and reveal stale or weakly indexed content.

A practical system prompt includes rules such as: distinguish direct evidence from inference, never invent owners or deadlines, prefer the newest approved source when sources conflict, and ask one focused clarification question when the request is ambiguous. Pass the model a compact evidence block containing chunk text and metadata rather than an unfiltered document dump. This lowers token cost and reduces distraction from irrelevant context.

Apply Permissions Before the Model Sees Content

Authorization belongs in retrieval, not merely in the interface. Filter every search query by the requesting user's groups and document access labels before sending results to the model. Never retrieve HR, legal, finance, customer, or security-restricted material and rely on the model to avoid mentioning it. Log document IDs and permissions decisions, but avoid retaining raw confidential prompts in analytics unless your retention policy explicitly permits it.

Also defend against prompt injection embedded in source files. Treat retrieved text as untrusted data. Delimit it clearly, tell the model that source content cannot alter system instructions, and strip or flag content that attempts to request secrets, tool calls, or policy overrides. For assistants with actions, separate answering from execution: show a proposed action, its parameters, and its source evidence before requiring explicit user confirmation.

Use Model Routing to Control Cost Without Weakening Quality

Most knowledge-assistant queries do not need a premium reasoning model. Route simple lookup, formatting, and query-rewrite tasks to a fast economical model, while reserving stronger models for multi-document synthesis, ambiguous policy interpretation, or complex troubleshooting. Cache embeddings, retrieval results for common questions, and approved answers keyed by corpus version.

59API is a practical option for this architecture because it provides low-cost, pay-as-you-go access to native official-quality Claude and GPT models through one relay. Its compatibility with OpenAI SDKs, Claude Code, and Codex means teams can usually change the API base URL to https://api.59api.com while retaining their existing client patterns. That makes it straightforward to test routing across Claude Opus, Sonnet, Haiku, Fable, and GPT options without building separate provider integrations. Its referral rebate can further reduce ongoing experimentation costs.

Evaluate Before You Expand

Create a versioned evaluation set from real employee questions, including expected sources and unacceptable answers. Score retrieval recall, citation correctness, factual completeness, refusal behavior, latency, and cost per resolved question. Review failures by category: missing content, bad chunking, access filtering, weak ranking, ambiguous wording, or model reasoning. Fix the earliest failing stage instead of repeatedly rewriting the final prompt.

Launch with one well-maintained domain, such as engineering runbooks or employee onboarding, and add a visible feedback control that captures whether the answer was useful, incorrect, outdated, or inaccessible. Assign owners to resolve recurring feedback and surface documents with no recent review. When you are ready to prototype a routed, cited assistant without committing to large minimum spend, sign up for 59API and validate the workflow against your team's real questions.