kb-rag

name	kb-rag
description	Knowledge Base RAG - retrieve and answer questions from indexed knowledge (blog translations, skills docs, project docs, memory archives). Use when user asks questions that might be answered by accumulated knowledge, wants to search across the knowledge base, or asks "我们知识库里有什么关于X的". Also use for ingesting new content into the knowledge base.
description_zh	知识库 RAG 检索：从已索引的知识（博客翻译、技能文档、项目文档、记忆归档）中检索和回答问题。

Knowledge Base RAG

Architecture

• Backend: qmd (BM25 + vector + hybrid reranker)
• LLM: Antigravity proxy (Gemini/Claude, $0 cost)
• Collections: 7 indexed collections, auto-updated every 5 min by OpenClaw

Collections

Collection	Content	Path
kb-articles	Blog translations & analysis	output/blog-translations/
kb-skills	Skill documentation	skills/**/SKILL.md
kb-captures	Conversation insight captures	content-pipeline/captures/
kb-docs	Project documentation	docs/
memory-dir	Daily notes & archives	memory/
memory-root	Core memory index	MEMORY.md
memory-alt	Alt memory	memory.md

Query (RAG Pipeline)

Run the RAG script for end-to-end retrieve + generate:

python3 skills/kb-rag/scripts/rag_query.py "question"

Options:

• -c <collection> — limit to specific collection (e.g. kb-articles)
• -k <N> — top-K chunks (default 8)
• -m <model> — LLM model (default gemini-3.1-pro-high)
• --retrieve-only — skip LLM, just show retrieved chunks
• --json — JSON output

Examples:

# General KB question
python3 skills/kb-rag/scripts/rag_query.py "OpenClaw 2026.2.13 有什么重要更新"

# Search only blog articles
python3 skills/kb-rag/scripts/rag_query.py "Agent 自主付费" -c kb-articles

# Just retrieve, no generation
python3 skills/kb-rag/scripts/rag_query.py "Peter Steinberger 的创业哲学" --retrieve-only

Manual Search (without LLM)

Direct qmd commands for quick lookups:

# Hybrid search (best quality)
qmd query "search terms" -n 5

# Vector similarity (better for semantic/fuzzy)
qmd vsearch "search terms" -n 5

# Full-text BM25 (exact term matching)
qmd search "search terms" -n 5

# Limit to collection
qmd query "search terms" -c kb-articles -n 5

Ingest New Content

Add individual files

New .md files dropped into indexed directories are auto-picked up on next qmd update && qmd embed.

Add new collection

qmd collection add /path/to/dir --name collection-name --mask "**/*.md"
qmd update && qmd embed

Manual re-index

qmd update    # re-index all collections
qmd embed     # regenerate embeddings

Note: OpenClaw runs qmd update every 5 minutes automatically. Run manually only when immediate indexing is needed.

Known Limitations

• Reranker (qwen3-reranker-0.6b) is weak on Chinese structured content — if results seem off, try qmd vsearch (pure vector) instead of qmd query (hybrid)
• Large files get chunked at 800 tokens with 15% overlap — very long articles may lose cross-chunk context