Semantic Code Exploration

Search code by meaning rather than keywords using semantic embeddings.

Workflow

1. Ensure Index Exists

Before searching, verify the codebase is indexed:

list_collections()

If the current working directory is not indexed:

index_codebase(directory="/absolute/path/to/project")

2. Search Semantically

Use complete behavioral questions, not keywords:

semantic_search(
  query="How does the system handle user authentication?",
  collection="collection-name"
)

To get only file locations without code chunks:

semantic_search(
  query="Where is authentication implemented?",
  collection="collection-name",
  include_code=False
)

3. Read Results

Review top results by score (>0.7 = strong match). Use file paths and line numbers to read relevant code.

Query Patterns

Instead of	Use
"auth handler"	"How does the system authenticate users?"
"error"	"What happens when a request fails?"
"database"	"How does the application connect to and query the database?"

Score Interpretation

• >0.7: Strong match - read this code
• 0.5-0.7: Moderate - review for relevance
• <0.5: Weak - refine query or use grep instead

When to Use Semantic vs Grep

Semantic Search	Grep/Glob
"How does X work?"	Known symbol names
Behavior/intent questions	Specific strings
Exploring unfamiliar code	File patterns

Collection Identification

Match collection to current project:

• Check list_collections() output
• Collection names often include project directory name or use ws-{hash} pattern
• Use get_collection_info(collection) to verify point count

Requirements

Services must be running:

• Qdrant: localhost:6333
• Embeddings: localhost:1335