Is multi-KB RAG a routing problem, not a retrieval problem?
Arc_bong · reddit · 2026-09-04
The author challenges the default "retrieve top-k from every source and fuse" enterprise RAG pattern:
- RRF fusion is fine with a handful of sources, but at 10+ KBs you're implicitly comparing scores from different retrieval distributions, domains and corpus sizes — cross-KB top-1 results get nearly identical fusion weight, and fixed similarity thresholds assume comparable score distributions that often don't exist
- Typical failure chain: good per-KB retrieval → questionable cross-KB ranking → bad context selection
He favors query → KB router selection → targeted retrieval → reranking → generation over retrieve-everywhere-then-fuse, at the cost of router errors and weaker coverage on genuinely cross-domain questions.
The insight came from comparing Lyzr Studio, LlamaIndex, LangChain and custom stacks; he asks what architectures actually hold up under dozens of knowledge sources and real production traffic.
Related event: Agentic RAG Debate: Route the Query, Don't Just Retrieve(2 posts)→
More from coding & agent
- Open-source Claude for SEO repo hits ~1k stars, cataloging 660+ agent-callable APIs — matchaman11 · 2026-09-05
- Claude for SEO: open-source tool pitches Claude as a free $1000/mo SEO suite, hits 1k stars — matchaman11 · 2026-09-05
- Tell your agent its budget: Shumer's tip for capping token spend per task — mattshumer_ · 2026-09-05
- Pokee Isaac connects to OpenAI Codex via the same MCP package built for Claude Code — Kyrannio · 2026-09-05
- Astra builds a working "macOS 27" with apps, terminal and browser in 75 minutes — altryne · 2026-09-05
- OpenWork launches Web version: a cloud computer with your full AI agent setup — ycombinator · 2026-09-05