I'm building a chatbot for employees that needs to answer from approved sources only. What architecture should I use?

Use a retrieval-augmented generation (RAG) architecture with strict source grounding and access control.

Recommended architecture

1. Approved source ingestion

   • Collect only sanctioned documents, wikis, policy pages, helpdesk articles, etc.
   • Normalize them into chunks with metadata:
     • source
     • title
     • URL/path
     • last updated
     • department
     • permission level

2. Indexing layer

   • Build a search index:
     • vector index for semantic retrieval
     • optionally keyword/BM25 for exact policy terms and names
   • Store metadata separately so you can filter by source/permission.

3. Query-time retrieval

   • Before generation, retrieve only from approved sources.
   • Apply:
     • document-level permissions
     • source allowlists
     • freshness filters
   • Retrieve top-k passages and pass them to the model.

4. Answer generation with citations

   • The LLM should answer only from retrieved context.
   • Require inline citations or footnotes to the exact source chunks.
   • If evidence is insufficient, the model must say:
     • “I don’t know based on the approved sources.”

5. Guardrails

   • Add a policy layer that:
     • blocks answers without citations
     • rejects unsupported claims
     • prevents use of external web data unless explicitly approved
   • Use a refusal template for out-of-scope or insufficient-evidence queries.

6. Audit and logging

   • Log:
     • query
     • retrieved documents
     • final answer
     • citations used
   • This helps with compliance and debugging.

Best practice pattern

A good pattern is:

User question → policy/permission check → retrieve from approved corpus → rerank → LLM answers with citations → post-check for groundedness → return or refuse

Key design choices

• Do not let the model browse freely unless that web content is also approved.
• Do not rely on the LLM’s memory for policy or employee info.
• Use metadata filters and ACLs so users only see what they’re allowed to see.
• Prefer “answer only with citations” for compliance-heavy environments.

If you want the simplest safe version

Start with:

• a single approved document store
• hybrid search
• a grounded-answer prompt
• citation requirement
• refusal if retrieval confidence is low

If you want, I can sketch a reference architecture diagram or propose a tech stack for this setup.