Replacing RAG pipelines with an autonomous agent that explores the knowledge base
Shivam__kumar · reddit · 2026-11-14
Instead of tuning retrieval logic per question type — where fixing one query often broke others — the author gave an autonomous agent (Claude Code) tools and a knowledge base and let it decide where to look and what steps to take.
- Tested against his traditional RAG implementation on an ERP system: surprisingly good results, hallucinations almost eliminated; remaining wrong answers mostly stemmed from gaps in his knowledge-base prep
- Open issues: knowledge base still needs maintenance; some lookups use keyword matching and predefined relations; each user currently needs a separate Claude Code process/session, and scaling to many users and much larger bases is unsolved
He's asking anyone with production RAG experience whether they've hit the same retrieval-maintenance pain.
More from coding & agent
- Restaurant owner built a $120 camera + vision model system that cut table bussing time in half — bennash · 2026-11-14
- Dev prediction: in the AI coding era, CS fundamentals beat framework knowledge — JFPuget · 2026-11-14
- GitHub testing Code Tour in Copilot app: code review as narrated chapters — 0xkarasy · 2026-11-14
- Small lab runs its whole dev stack with Claude Skills and subagents — seanwbren · 2026-11-14
- Keyfleet to open fleet creation with group access rules and personal persistent-agent teams — seanwbren · 2026-11-14
- Regents Labs launches 6 agent-native apps with token-owned revenue split in USDC on Base — seanwbren · 2026-11-14