Enterprise Knowledge Agent

A Q&A assistant over internal company docs (Confluence, Drive, Jira, Slack) that enforces per-user permissions - users only get answers from docs they're allowed to read. The agent ingests documents from multiple sources, stores embeddings alongside ACL metadata, enforces pre-filter permission checks at query time, retrieves with hybrid search, and synthesizes grounded answers with source citations.

“Permission enforcement is non-negotiable and must be pre-filter. ACL lives with the embedding in the index - not applied after retrieval. Every query is bound to an authenticated identity. Agent says I don't know rather than synthesize beyond permitted corpus.”

[Confluence / Drive / Jira / Slack]
        ↓ (connectors, webhooks)
[1] Ingest → Parse → Chunk → Embed → Store with ACL metadata
        ↓
[2] Permission-Aware Vector Index  ←── Identity Provider (Okta / Google Workspace)
        ↓
User Query → [3] Auth + Identity Resolution → Audit log entry created
        ↓
[4] Permission Enforcement  (pre-filter: only permitted doc IDs enter search)
        ↓
[5] Hybrid Retrieval  (vector + BM25) → Re-rank
        ↓
[6] Context Assembly  (top-N chunks + source labels)
        ↓
[7] LLM Synthesis  (grounded answer + citations)
        ↓
Answer → User  +  Audit log completed

• 1Ingestion Pipeline - One connector per source (Confluence API, Drive API, Jira REST). Track last_synced_at per doc for incremental sync. Chunk at ~512 tokens with 50-token overlap. Store { doc_id, url, acl_list, last_modified } alongside each vector.
• 2Permission-Aware Index - Most Critical Component - At index time, fetch the doc's ACL from the source and store it with the embedding. Cache user_id → permitted_doc_ids with 15-min TTL. Invalidate on permission-change events from IdP.
• 3Permission Enforcement - Pre-filter vs Post-filter - Pre-filter (recommended): pass ACL filter to vector DB query - unauthorized docs never fetched. Post-filter is unacceptable in regulated environments because unauthorized docs enter memory. Use pre-filter always.
• 4Hybrid Retrieval - Dense vector (semantic) + sparse BM25 (keyword). Enterprise queries have exact terms (project codenames, ticket IDs) that pure vector search misses. Merge with Reciprocal Rank Fusion (RRF), then apply cross-encoder re-ranker.
• 5LLM Synthesis - Prompt: answer from context only, cite every claim with [Doc Title](url), say 'I don't know' if answer isn't in context. PII redaction pass before sending chunks to cloud LLM.
• 6Audit Log - Immutable, append-only (BigQuery / S3 Object Lock). Record: { user_id, query, sources_cited, answer_summary, timestamp }. Compliance dashboard: who asked what, which docs were surfaced.

• 1Permission staleness - Doc goes private but served for 15 min → mitigated by event-driven invalidation from IdP on permission changes.
• 2Corpus quality - Wikis are outdated and contradictory → freshness boost on retrieval + 'I don't know' grounding when no relevant content exists.
• 3LLM grounding drift - Model synthesizes beyond retrieved context → strict context-only prompt instruction + mandatory source citation requirement.

Google Stack: Google Workspace (source) → Vertex AI Search with data store ACL integration → Gemini (synthesis LLM) → Vertex AI Vector Search (permission-filtered retrieval) → BigQuery (audit log) → Cloud Identity / Google Workspace IdP (identity resolution)

Azure Stack: SharePoint / OneDrive / Azure DevOps (sources) → Azure Data Factory (connectors) → Azure AI Search / Foundry IQ (hybrid BM25 + vector, native ACL pre-filter) → Azure Entra ID (identity + transitive group resolution) → Azure OpenAI GPT-4o (synthesis LLM) → Azure Cache for Redis (permission cache, 15-min TTL) → Azure Monitor (audit log)

AWS Stack: Confluence / SharePoint / S3 (sources) → AWS Glue (connectors) → Amazon Bedrock Knowledge Bases + OpenSearch Serverless (managed RAG, hybrid retrieval) → AWS IAM Identity Center (identity resolution) → Amazon Bedrock / Claude (synthesis LLM) → ElastiCache Redis (permission cache) → CloudTrail + CloudWatch Logs (audit log)

Other Options: Elasticsearch + Qdrant + OpenAI (open-source, full control over ACL logic) | Pinecone + LlamaIndex + Anthropic Claude (serverless vector DB path) | Microsoft 365 Copilot (if Microsoft-only stack, zero-integration, but limited customization)