Semantic Project Indexing

Project Indexing provides Sypha with an advanced semantic search capability. Unlike traditional keyword matches, Sypha interprets the conceptual meaning behind your prompts, allowing it to locate relevant structures and logic even when exact terms aren't known.

Operational Workflow

1. Structural Analysis: Sypha utilizes Tree-sitter to parse your source code into logical components like classes and methods.
2. Vector Generation: These components are processed through an AI model to create numerical vector embeddings.
3. Optimized Storage: The vectors are managed within a Qdrant database, facilitating high-speed similarity retrieval.
4. Context Discovery: Sypha leverages the codebase_search utility to inject relevant project context into active tasks.

Infrastructure Setup

1. Intelligence Provider

Select your preferred service for generating embeddings:

• Cloud Standard: High-fidelity processing via text-embedding-3-small (OpenAI).
• Scale Optimized: A balanced approach using Google's Gemini models.
• Private/Local: Process everything on-device using Ollama for maximum confidentiality.

2. Retrieval Engine (Qdrant)

A running Qdrant instance is required to host your index:

• Containerized: Deploy instantly via Docker: docker run -p 6333:6333 qdrant/qdrant.
• Managed: Utilize Qdrant Cloud for hosted infrastructure.

Internal Management

Adjust your indexing preferences in Settings (gear icon) -> Codebase Indexing.

• Service Toggle: Activate or pause the background indexing process as needed.
• Lifecycle States:
  • Standby: The indexing engine is currently inactive.
  • Indexing: Actively processing project updates.
  • Indexed: The project map is synchronized and available for querying.

Strategic Tips

• Broad Compatibility: Designed for languages with robust Tree-sitter parsers.
• Confidentiality: Analysis is performed locally; only abstract numerical representations are transmitted to cloud providers.
• Smart Exclusions: Sypha automatically bypasses node_modules, dist, and other ignored paths to maintain a clean, relevant index.