AI 新闻雷达
返回
RAG 与知识库·github.com

bug(rag): query side appends /faiss_index but ingest writes to vectordb_dir directly → KB retrieval never finds the index (#1951 follow-on)

目标用户 · developer

阅读原文

bug(rag): query side appends /faiss_index but ingest writes to vectordb_dir directly → KB retrieval never finds the index (#1951 follow-on): ## Severity **High** — RAG retrieval is broken for the standard config path: `query_knowledge_base` returns *"no knowledge base found"* even when an index was successfully ingested. Affects any setup where ingest and que

个性化解读(发生了什么 / 为什么重要 / 影响 / 建议)将在接入 LLM 后按你的角色生成。

痛点信号

  • bug(rag): query side appends /faiss_index but ingest writes to vectordb_dir directly → kb retrieval never finds the index (#1951 follow-on).
  • then `_collect_faiss_dirs()` (`src/tools/rag.py`) does `_has_faiss_index( )` → checks `/data/vectordb/faiss_index/index.faiss`, which **doesn't exist** → no dirs collected → *"no knowledge base found."* **so: #1951 removed the `/faiss_index` append from the ingest side but left it in `configure_rag_tool` on the query side.** this inverts the original doubling bug into a one-sided mismatch, and `configure_rag_tool` now contradicts the convention the `ingestconfig` docstring says it follows.
  • found while load-testing a hardened container config; the index built fine and memory was well within limits, but retrieval returned nothing — traced to this path asymmetry.
#developer#rag#issues