using-agent-brain（使用 Agent 大脑）

Agent Brain Expert Skill

Expert-level skill for Agent Brain document search with five modes: BM25 (keyword), Vector (semantic), Hybrid (fusion), Graph (knowledge graph), and Multi (comprehensive fusion).

Contents

• Search Modes
• Mode Selection Guide
• GraphRAG (Knowledge Graph)
• Indexing & Folder Management
• Content Injection
• Job Queue Management
• Server Management
• Cache Management
• When Not to Use
• Best Practices
• Reference Documentation

Search Modes

Mode Parameters

Mode Selection Guide

Use BM25 When

Searching for exact technical terms:

agent-brain query "recursiveCharacterTextSplitter" --mode bm25
agent-brain query "ValueError: invalid token" --mode bm25
agent-brain query "def process_payment" --mode bm25

Counter-example - Wrong mode choice:

# BM25 is wrong for conceptual queries
agent-brain query "how does error handling work" --mode bm25  # Wrong
agent-brain query "how does error handling work" --mode vector  # Correct

Use Vector When

Searching for concepts or natural language:

agent-brain query "best practices for error handling" --mode vector
agent-brain query "how to implement caching" --mode vector

Counter-example - Wrong mode choice:

# Vector is wrong for exact function names
agent-brain query "getUserById" --mode vector  # Wrong - may miss exact match
agent-brain query "getUserById" --mode bm25    # Correct - finds exact match

Use Hybrid When

Need comprehensive results (default mode):

agent-brain query "OAuth implementation" --mode hybrid --alpha 0.6
agent-brain query "database connection pooling" --mode hybrid

Alpha tuning:

• --alpha 0.3 - More keyword weight (technical docs)
• --alpha 0.7 - More semantic weight (conceptual docs)

Use Graph When

Exploring relationships and dependencies:

agent-brain query "what functions call process_payment" --mode graph
agent-brain query "classes that inherit from BaseService" --mode graph --traversal-depth 3
agent-brain query "modules that import authentication" --mode graph

Prerequisite: Requires ENABLE_GRAPH_INDEX=true during server startup.

Use Multi When

Need the most comprehensive results:

agent-brain query "complete payment flow implementation" --mode multi --include-relationships

GraphRAG (Knowledge Graph)

GraphRAG enables relationship-aware retrieval by building a knowledge graph from indexed documents.

Enabling GraphRAG

export ENABLE_GRAPH_INDEX=true
agent-brain start

Graph Query Types

See Graph Search Guide for detailed usage.

Indexing & Folder Management

Indexing with File Type Presets

# Index only Python files
agent-brain index ./src --include-type python

# Index Python and documentation
agent-brain index ./project --include-type python,docs

# Index all code files
agent-brain index ./repo --include-type code

# Force full re-index (bypass incremental)
agent-brain index ./docs --force

Use agent-brain types list to see all 14 available presets.

Folder Management

agent-brain folders list                    # List indexed folders with chunk counts
agent-brain folders add ./docs              # Add folder (triggers indexing)
agent-brain folders add ./src --include-type python  # Add with preset filter
agent-brain folders remove ./old-docs --yes # Remove folder and evict chunks

Incremental Indexing

Re-indexing a folder automatically detects changes:

• Unchanged files are skipped (mtime + SHA-256 checksum)
• Changed files have old chunks evicted and new ones created
• Deleted files have their chunks automatically removed
• Use --force to bypass manifest and fully re-index

Content Injection

Enrich chunk metadata during indexing with custom Python scripts or static JSON metadata.

When to Use

• Tag chunks with project/team/category metadata
• Classify chunks by content type
• Add custom fields for filtered search
• Merge folder-level metadata into all chunks

Basic Usage

# Inject via Python script
agent-brain inject ./docs --script enrich.py

# Inject via static JSON metadata
agent-brain inject ./src --folder-metadata project-meta.json

# Validate script before indexing
agent-brain inject ./docs --script enrich.py --dry-run

Injector Script Protocol

Scripts export a process_chunk(chunk: dict) -> dict function:

def process_chunk(chunk: dict) -> dict:
    chunk["project"] = "my-project"
    chunk["team"] = "backend"
    return chunk

• Values must be scalars (str, int, float, bool)
• Per-chunk exceptions are logged as warnings, not fatal
• See docs/INJECTOR_PROTOCOL.md for the full specification

Job Queue Management

Indexing runs asynchronously via a job queue. Monitor and manage jobs:

agent-brain jobs                    # List all jobs
agent-brain jobs --watch            # Live polling every 3s
agent-brain jobs <job_id>           # Job details + eviction summary
agent-brain jobs <job_id> --cancel  # Cancel a job

Eviction Summary

When re-indexing, job details show what changed:

Eviction Summary:
  Files added:     3
  Files changed:   2
  Files deleted:   1
  Files unchanged: 42
  Chunks evicted:  15
  Chunks created:  25

This confirms incremental indexing is working efficiently.

Server Management

Quick Start

agent-brain init              # Initialize project (first time)
agent-brain start    # Start server
agent-brain index ./docs      # Index documents
agent-brain query "search"    # Search
agent-brain stop              # Stop when done

Progress Checklist:

• /agent-brain:agent-brain-init succeeded
• /agent-brain:agent-brain-status shows healthy
• Document count > 0
• Query returns results (or "no matches" - not error)

Lifecycle Commands

Pre-Query Validation

Before querying, verify setup:

agent-brain status

Expected:

• Status: healthy
• Documents: > 0
• Provider: configured

Counter-example - Querying without validation:

# Wrong - querying without checking status
agent-brain query "search term"  # May fail if server not running

# Correct - validate first
agent-brain status && agent-brain query "search term"

See Server Discovery Guide for multi-instance details.

Cache Management

The embedding cache automatically stores computed embeddings to avoid redundant API calls during reindexing. No setup is required — the cache is active by default.

When to Check Cache Status

• After indexing — verify cache is working and hit rate is growing
• When queries seem slow — a low or zero hit rate means embeddings are being recomputed on every reindex
• To monitor cache growth — track disk usage over time for large indexes

agent-brain cache status

A healthy cache shows:

• Hit rate > 80% after the first full reindex cycle
• Growing disk entries over time as more content is indexed
• Low misses relative to hits

When to Clear the Cache

• After changing embedding provider or model — prevents dimension mismatches and stale cached vectors
• Suspected cache corruption — if embeddings seem incorrect or search quality degrades unexpectedly
• To force fresh embeddings — when you need to ensure all vectors reflect the current provider/model

# Clear with confirmation prompt
agent-brain cache clear

# Clear without prompt (use in scripts)
agent-brain cache clear --yes

Cache is Automatic

No configuration is required. Embeddings are cached on first compute and reused on subsequent reindexes of unchanged content (identified by SHA-256 hash). The cache complements the ManifestTracker — files that haven't changed on disk won't need to recompute embeddings.

See the API Reference for GET /index/cache and DELETE /index/cache endpoint details, including response schemas.

When Not to Use

This skill focuses on searching and querying. Do NOT use for:

• Installation - Use configuring-agent-brain skill
• API key configuration - Use configuring-agent-brain skill
• Server setup issues - Use configuring-agent-brain skill
• Provider configuration - Use configuring-agent-brain skill

Scope boundary: This skill assumes Agent Brain is already installed, configured, and the server is running with indexed documents.

Best Practices

1. Mode Selection: BM25 for exact terms, Vector for concepts, Hybrid for comprehensive, Graph for relationships
2. Threshold Tuning: Start at 0.7, lower to 0.3-0.5 for more results
3. Server Discovery: Use runtime.json rather than assuming port 8000
4. Resource Cleanup: Run agent-brain stop when done
5. Source Citation: Always reference source filenames in responses
6. Graph Queries: Use graph mode for "what calls X", "what imports Y" patterns
7. Traversal Depth: Start with depth 2, increase to 3-4 for deeper chains
8. File Type Presets: Use --include-type python,docs instead of manual glob patterns
9. Incremental Indexing: Re-index without --force for efficient updates
10. Injection Validation: Always --dry-run injector scripts before full indexing
11. Job Monitoring: Use agent-brain jobs --watch for long-running index jobs

Reference Documentation

Limitations

• Vector/hybrid/graph/multi modes require embedding provider configured
• Graph mode requires additional memory (~500MB extra)
• Supported formats: Markdown, PDF, plain text, code files (Python, JS, TS, Java, Go, Rust, C, C++)
• Not supported: Word docs (.docx), images
• Server requires ~500MB RAM for typical collections (~1GB with graph)
• Ollama requires local installation and model download