graph-search

---

name: graph-search description: Hybrid search combining vector similarity with graph traversal for multi-hop reasoning over the knowledge graph disable-model-invocation: false argument-hint: [query] [--project project_name] allowed-tools: Bash, Read

Graph Search

Hybrid search that combines vector similarity with graph traversal to discover and synthesize knowledge across the full graph structure.

Usage

/graph-search "What technologies are used in the project?"
/graph-search "generative AI projects" --project project_a

How it differs from /vector-search

• /vector-search: Returns top-K similar nodes (isolated matches)
• /graph-search: Finds seed nodes via vector search, then traverses RELATES_TO, BELONGS_TO, MENTIONS, and SOURCED_FROM to gather multi-hop context, then synthesizes a coherent answer

Workflow

1. Run Graph Search

python ${CLAUDE_SKILL_DIR}/../../../scripts/graph_search.py "$QUERY" $1 $2 --json

This returns JSON with:

• seed_entities: Top entities by vector similarity
• seed_chunks: Top chunks by vector similarity
• seed_communities: Top communities by vector similarity
• related_entities: Entities connected via RELATES_TO (graph expansion)
• entity_communities: Communities the entities belong to
• context_chunks: Source text chunks mentioning relevant entities
• provenance: Which documents contributed each entity

2. Read Context Chunks

If the context chunks from Step 1 are truncated or insufficient, read the full text of the most relevant chunks. Use the source_path from provenance to read original documents if needed.

3. Synthesize Answer

Using all gathered context (seed nodes, related entities, communities, chunks, provenance), synthesize a comprehensive answer to the user's question:

• Cite specific entities and their relationships
• Reference source documents via provenance
• Use community context to provide broader thematic understanding
• If the graph lacks sufficient information, say so clearly rather than speculating

4. Show Sources

At the end, list the source documents and key entities that informed the answer.

Options

• --top / -k: Number of seed nodes per type (default: 3)
• --max-related: Max related entities to expand (default: 10)
• --max-chunks: Max context chunks (default: 10)
• --project / -p: Project name