feat: add Tavily as a pluggable internet search backend

[ParakhJaggi]

Summary

• Added Tavily as a fifth pluggable backend in the InternetRetrieverFactory system, keeping BochaAI as the default
• Created TavilySearchRetriever class using the tavily-python SDK, following the same retrieve_from_internet interface and TextualMemoryItem metadata schema as BochaAISearchRetriever
• Registered TavilySearchConfig in InternetRetrieverConfigFactory.backend_to_class and wired instantiation in InternetRetrieverFactory.from_config()
• Updated APIConfig.get_internet_config() to auto-select Tavily when TAVILY_API_KEY is set and BOCHA_API_KEY is absent
• Added tavily-python as an optional dependency under a new tavily extra and in the all group
• Added a Step 7 Tavily demo block to the internet search example, gated on TAVILY_API_KEY

Files changed

• src/memos/memories/textual/tree_text_memory/retrieve/tavilysearch.py (new)
• src/memos/configs/internet_retriever.py
• src/memos/memories/textual/tree_text_memory/retrieve/internet_retriever_factory.py
• src/memos/api/config.py
• pyproject.toml
• examples/basic_modules/textual_memory_internet_search_example.py

Dependency changes

• Added tavily-python (>=0.5.0,<1.0.0) as optional dependency under new tavily extra
• Added tavily-python to the all extras group

Environment variable changes

• Added TAVILY_API_KEY — API key for Tavily Search (additive; existing BOCHA_API_KEY, GOOGLE_API_KEY, GOOGLE_SEARCH_ENGINE_ID unchanged)

Notes for reviewers

• This is an additive change — no existing backends or code paths are modified
• Backend selection priority: if TAVILY_API_KEY is set and BOCHA_API_KEY is not set, Tavily is used; otherwise BochaAI remains the default
• The Tavily retriever uses search_depth="basic" for cost efficiency; can be changed to "advanced" for higher relevance

Automated Review

• Passed after 1 attempt(s)
• Final review: The Tavily migration is well-implemented and follows existing patterns correctly. All six files from the plan are addressed. The TavilySearchRetriever correctly uses the tavily-python SDK (TavilyClient, response["results"] field access), mirrors the BochaAI retriever interface (retrieve_from_internet signature, TextualMemoryItem construction, metadata schema), and uses require_python_package for lazy dependency guarding. Config registration, factory wiring, dependency declarations, and the example are all properly done. Three minor issues were found — none block approval.