v0.1.1-alpha - Smart Context & Hybrid Search

🚀 Summary

A major upgrade to how agents understand and search codebases, introducing hybrid retrieval and token-aware context gathering.

✨ Key Features & Improvements

🔍 Hybrid Knowledge Retrieval

• What changed: Integrated Qdrant for both Dense (semantic) and Sparse (keyword) vector embeddings.
• Reasoning: Semantic search alone can miss specific technical terms (like variable names). Hybrid search combines the conceptual understanding of LLMs with the precision of keyword matching.

🧠 Smart Context Gathering

• What changed: Implemented a "budget-aware" context builder that intelligently gathers relevant files, PR diffs, and project history.
• Reasoning: LLMs have limited context windows. By prioritizing the most relevant "smart" context, we reduce hallucination and token cost while improving solution quality.

🌎 Language-Aware Reviewers

• What changed: Added filtering logic that selects review agents based on the file types modified in a PR.
• Reasoning: Prevents "noise" and wasted tokens by ensuring a Python Security specialist doesn't review CSS changes.

🐛 Issues Addressed

• #8 [Data Integrity]: Initial implementation of vector embeddings for more reliable knowledge lookup.
• #11 [Architecture]: Enhanced smart context gathering for cleaner agent inputs.

📝 Technical Notes

• Added qdrant-client and fastembed dependencies.
• Migrated codebase analysis to use recursive AST parsing for better context depth.