…xpansion

- Remove _should_use_deep_search and find() auto-selection
- Always use search() with session for IntentAnalyzer
- Add _expand_query() to expand Chinese queries with English keywords
- Example: 'OpenViking 怎么安装' → 'OpenViking 怎么安装 installation setup install pip'
- Installation doc now ranks #1 (was not in top 5 before)

* feat: add Memex personal knowledge assistant demo

- CLI-based personal knowledge assistant powered by OpenViking
- Knowledge management: add files, directories, URLs
- Intelligent Q&A with RAG and multi-turn conversation
- Knowledge browsing with L0/L1/L2 context layers
- Semantic search with intent analysis
- Feishu/Lark MCP integration for document import
- Rich CLI interface with prompt_toolkit

* fix: fix Memex bugs in tree display and RAG config loading

- Fix /tree command to handle dict list return type
- Fix RAG recipe to load LLM config from ov.conf instead of removed attributes
- Fix find/search methods to use 'limit' parameter instead of 'top_k'

* fix: fix URI path handling in abstract/overview commands

- Strip trailing slashes from URIs to prevent path errors
- Improve tree display to handle dict list return type

* feat: improve Memex directory support and cleanup

- Add recursive directory import with file type filtering
- Improve error handling for add_resource results
- Add proper cleanup on exit to kill AGFS processes
- Fix import_dir to use new directory handling logic

* fix: improve RAG retrieval by lowering score threshold and reading full content

- Lower default score_threshold from 0.3 to 0.1 (actual scores are 0.17-0.25)
- Fix build_context to use client.read() for full content instead of just abstract
- This fixes the issue where queries returned no results despite data being indexed

* feat: add Tier 1 RAG improvements - query rewriting, confidence hints, smart context loading

* feat: add Tier 2 RAG improvements - Session integration, memory extraction, hybrid search, rerank

- P0: Integrate OpenViking Session for context-aware search with IntentAnalyzer
- P1: Add session.commit() for long-term memory extraction (6 categories)
- P2: Implement hybrid search strategy (find for simple, search for complex)
- P3: Add LLM-based rerank for improved result relevance

New CLI commands: /session, /commit, /memories

* fix: fix content retrieval in _process_search_results

- abstract() only works for directories, not files
- Changed to try read() first, then abstract() for directories
- This fixes the issue where only 1 result was returned instead of 5

* refactor: simplify RAG recipe - remove dual thresholds and rerank

- Remove high_confidence_threshold and low_confidence_threshold
- Remove rerank (use OpenViking score directly)
- Remove confidence hints from system prompt
- Unified content loading: all results get full content (2000 chars)
- Simpler build_context without confidence logic

* feat: improve intent understanding - always use deep search + query expansion

- Remove _should_use_deep_search and find() auto-selection
- Always use search() with session for IntentAnalyzer
- Add _expand_query() to expand Chinese queries with English keywords
- Example: 'OpenViking 怎么安装' → 'OpenViking 怎么安装 installation setup install pip'
- Installation doc now ranks #1 (was not in top 5 before)

* docs: update memex README with architecture details, remove redundant query expansion

* fix: address PR review feedback

- Add wait_processed() after session commit for memory embedding
- Add OpenAI api_base default (https://api.openai.com/v1)
- Remove [:2000] content truncation that caused unexpected results

…or local dev (#237)

* feat: add Memex personal knowledge assistant demo

- CLI-based personal knowledge assistant powered by OpenViking
- Knowledge management: add files, directories, URLs
- Intelligent Q&A with RAG and multi-turn conversation
- Knowledge browsing with L0/L1/L2 context layers
- Semantic search with intent analysis
- Feishu/Lark MCP integration for document import
- Rich CLI interface with prompt_toolkit

* fix: fix Memex bugs in tree display and RAG config loading

- Fix /tree command to handle dict list return type
- Fix RAG recipe to load LLM config from ov.conf instead of removed attributes
- Fix find/search methods to use 'limit' parameter instead of 'top_k'

* fix: fix URI path handling in abstract/overview commands

- Strip trailing slashes from URIs to prevent path errors
- Improve tree display to handle dict list return type

* feat: improve Memex directory support and cleanup

- Add recursive directory import with file type filtering
- Improve error handling for add_resource results
- Add proper cleanup on exit to kill AGFS processes
- Fix import_dir to use new directory handling logic

* fix: improve RAG retrieval by lowering score threshold and reading full content

- Lower default score_threshold from 0.3 to 0.1 (actual scores are 0.17-0.25)
- Fix build_context to use client.read() for full content instead of just abstract
- This fixes the issue where queries returned no results despite data being indexed

* feat: add Tier 1 RAG improvements - query rewriting, confidence hints, smart context loading

* feat: add Tier 2 RAG improvements - Session integration, memory extraction, hybrid search, rerank

- P0: Integrate OpenViking Session for context-aware search with IntentAnalyzer
- P1: Add session.commit() for long-term memory extraction (6 categories)
- P2: Implement hybrid search strategy (find for simple, search for complex)
- P3: Add LLM-based rerank for improved result relevance

New CLI commands: /session, /commit, /memories

* fix: fix content retrieval in _process_search_results

- abstract() only works for directories, not files
- Changed to try read() first, then abstract() for directories
- This fixes the issue where only 1 result was returned instead of 5

* refactor: simplify RAG recipe - remove dual thresholds and rerank

- Remove high_confidence_threshold and low_confidence_threshold
- Remove rerank (use OpenViking score directly)
- Remove confidence hints from system prompt
- Unified content loading: all results get full content (2000 chars)
- Simpler build_context without confidence logic

* feat: improve intent understanding - always use deep search + query expansion

- Remove _should_use_deep_search and find() auto-selection
- Always use search() with session for IntentAnalyzer
- Add _expand_query() to expand Chinese queries with English keywords
- Example: 'OpenViking 怎么安装' → 'OpenViking 怎么安装 installation setup install pip'
- Installation doc now ranks #1 (was not in top 5 before)

* docs: update memex README with architecture details, remove redundant query expansion

* fix: address PR review feedback

- Add wait_processed() after session commit for memory embedding
- Add OpenAI api_base default (https://api.openai.com/v1)
- Remove [:2000] content truncation that caused unexpected results

* feat: add --verbose flag, fix README quick start

- Add -v/--verbose flag to control logging (default: off)
- Fix Quick Start: add cd examples/memex step
- Update CLI Options to match actual parameters

* fix(memex): add build-system and use relative imports so 'uv run memex' works

pyproject.toml was missing [build-system], causing uv to skip entry point
installation. Also fixed entry point from 'cli:main' to 'memex.cli:main'
and converted all bare imports to relative imports. Updated README to
match actual project structure.

* fix(memex): suppress openviking INFO logs in non-verbose mode

openviking creates per-module loggers with their own handlers and
propagate=False, making root logger level ineffective. Patch get_logger
to enforce WARNING level for all current and future openviking loggers.

* feat(memex): add full Feishu OAuth support and fix tool invocation

- Implement OAuth 2.0 authorization code flow for user access token
- Add support for manual code entry (for headless/SSH environments)
- Fix tool names (underscore format) and parameter structure (nested)
- Replace MCP file listing with REST API search (bypassing MCP limitations)
- Add /feishu-login and improve /feishu-list commands

* feat(memex): add /feishu-ls command for directory browsing

- Implement /feishu-ls using drive.v1.file.list API
- Allows browsing user's 'My Space' (root) and subfolders
- Requires user access token via /feishu-login
- Updated docs with redirect URI configuration step

* fix(memex): auto-resolve root folder token when folder_token is empty

* fix(memex): adapt memex for local dev with OpenAI provider

- Fix VLM backend config fallback (provider field support)
- Use find() instead of search() to avoid IntentAnalyzer deadlock
- Add uv source override for local openviking editable install
- Replace deleted logger monkey-patch with standard logging
- Update README: add missing commands, OpenAI config example, fix structure

* refactor(memex): switch back to client.search() for multi-type retrieval

* fix(async): make run_async re-entrant safe to prevent deadlock

Session methods call run_async() from within code already running on
the shared background event loop, causing a deadlock when search() is
used with session context. Fix by detecting a running event loop via
asyncio.get_running_loop() and spawning a temporary thread+loop for
the nested coroutine.