Vector search was failing with "unable to retrieve context" because file content wasn't being retrieved from the database and passed to the LLM, making RAG ineffective.

Core Fix

• Vector search now retrieves file content: search_semantic() calls get_chunk_text() to fetch actual code chunks, not just metadata
• Content always included: Removed include_content flag (de facto always true for RAG to work)
• llama-index integration: Replaced custom embedding/chunking with industry-standard framework

llama-index Migration

Created three new components:

• OpenAICompatibleEmbedding - wraps OpenAI API for llama-index
• chunk_with_llama_index - uses CodeSplitter (code-aware) and SentenceSplitter
• SQLiteVectorStore - bridges llama-index VectorStore interface with sqlite-vector backend

Before:

# Custom implementation, content not retrieved
results = _search_vectors(database_path, q_emb, top_k=top_k)
# Returns: [{file_id, path, chunk_index, score}]

After:

# llama-index with content retrieval
docs = llama_index_search(query, database_path, top_k=top_k)
results = [{..., "content": doc.text}]  # Content always included

UI Enhancements

• Markdown rendering: marked.js + highlight.js for code syntax highlighting
• Progress tracking: "Indexed: 536/1200 files | 6594 chunks" (total files stored in metadata, not recalculated)
• Incremental indexing: Continue (+) vs Full Re-index (⟳) buttons
• Custom projects: Form to add/index projects from any path
• File visibility: Shows which files were used in response (paths + scores)

Performance

• Total files calculated once during indexing, stored in metadata (was: directory walk on every request)
• Incremental indexing skips unchanged files based on mtime + hash
• Batch metadata updates in single transaction

Cleanup

Removed ~600 lines of obsolete code:

• EmbeddingClient class (replaced by llama-index)
• smart_chunker.py (replaced by CodeSplitter/SentenceSplitter)
• chunk_text utility (replaced by llama-index)
• include_content parameter (always needed)

Backwards Compatibility

No database schema changes. Existing projects continue to work. New indexes store total_files metadata for performance.

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.