- Replace description_text_full with description_text in vector search (reduce token bloat)
- Add description_text to CASE 1 (code search) and CASE 3 (vector search)
- Fix generic word fallback: 'quần bán chạy' now filters correctly
- Derive GENERIC_WORDS from PRODUCT_LINE_MAP (auto-sync)
- Fix duplicate if include_product_ids in conversation_manager.py
- Add product_mapping.py with PRODUCT_LINE_MAP and resolve functions

- Add master_color as explicit SQL FILTER field in SearchItem (was in description)
- Add LIKE filter in _get_metadata_clauses for master_color + product_color_name
- Add MAX_BY(master_color) to final SELECT aggregation
- Fix color fallback in format_product_results (parse from description_text_full)
- Strengthen Rule 3a: specific design/pattern matching (con lợn, Disney, etc.)
- Add thảo mai exception: truthfulness > flattery when no products match
- Update tool prompt examples to use master_color as separate field

fix: accessory filter mapping - use actual DB product_line_vn values for accessories (Khan, Mu, Tat, Tui xach, Chan ca nhan etc.) + remove 8-product truncation limit + add accessory guidance in prompts

fix: disable strict filters (sleeve, style, fitting, neckline, material, season) - only keep name, gender, age, color

feat: optimize filter performance and rename query to description - Pre-compile regex patterns for filter optimization - Rename SearchItem.query to SearchItem.description for semantic search - Add váy đầm, váy bé gái to product_map - Fix price format (remove decimal) - Tune LIMIT values for vector search

Update prompt: Enhance [NEXT] strategy, improve writing style, refine [GOAL] handling, and inject insight into prompt

feat: optimize logging, enable embedding cache, fix rate limiting hanging, and improve pagination security

fix: ensure price fields always have numeric values - Add default 0 for price/sale_price to prevent FE toLocaleString errors - Add debug logging in controller and data_retrieval_tool

feat: refactor product data retrieval - Fix URL parsing in description_text_full - Update system_prompt to require product_ids in responses - Remove graph singleton for hot reload support - Improve product enrichment logic

- Add reset_limit.py service for tracking daily reset usage via Redis
- Add archive_history method in conversation_manager.py to rename old messages
- Add POST /api/history/archive endpoint with limit check
- Update frontend with Reset button next to Send button
- Limit LLM context to 15 messages per session
- Filter chat history by current date (auto-clear next day)

- Add create_embeddings_async() to support OpenAI batch embedding API
- Refactor data_retrieval_tool to batch embed all queries in ONE request
- Replace print() with logger.info() in product_search_helpers
- Remove visual_search checks (only text search supported)

Performance: 5-10x faster for multi-search queries (300ms vs 1.5s for 5 queries)
Rate Limit: Saves RPM by batching multiple embeddings into single API call

- Dockerfile.dev: Development environment with hot reload & cache mount
- Add Makefile targets: build-dev, run-dev for easy local testing
- Hot reload watches agent, api, common directories
- Includes health check endpoint

Usage:
  make build-dev  # Build dev image
  make run-dev    # Run container with volume mount for hot reload

- Both Dockerfile.stage and Dockerfile.prod now use --mount=type=cache
- Faster builds with BuildKit cache mount (temporary, per-build)
- Requires: docker buildx or DOCKER_BUILDKIT=1

- Remove --no-cache-dir flag from pip install
- Use Docker layer caching for faster production builds
- Manual server rebuilds will auto-update packages if needed