Tools

This directory defines the public HybridDeepResearch tool contract. It is the environment an agent is expected to talk to, not a complete agent runtime.

The names, arguments, and descriptions live in schemas.json and can be loaded as OpenAI-style function tools:

from tools import openai_tools, build_toolset

tools = openai_tools()
callables = build_toolset(
    sqlite_db_path="/path/to/cross_db/cross_db.sqlite",
    sqlite_root="/path/to/sqlite_databases",
    db="cross_db",
)

Required capabilities

Tool	Purpose
`web_search(query, count)`	Discover sources and URLs
`web_fetch(url)`	Read a selected page
`run_sql(sql, description)`	Execute a read-only query
`final_answer(answer)`	End the run with one answer

Optional knowledge tools

When a database ships a <db>_kb.jsonl file next to its SQLite file, the agent may also use:

Tool	Purpose
`get_all_external_knowledge_names()`	List domain-knowledge titles
`get_knowledge_definition(knowledge_name)`	Load one definition
`get_all_knowledge_definitions()`	Load the full catalog, or titles if it is too large

Rules

Database access is read-only. run_sql accepts SELECT / WITH / PRAGMA / EXPLAIN and rejects writes. The SQLite connection is opened in mode=ro.
web_search / web_fetch here are reference backends. The default search path uses TAVILY_API_KEY if you set it; otherwise pass your own search_fn. Participants may replace the search and fetch implementations as long as the tool names and argument shapes stay compatible if they want drop-in reuse.
Call final_answer exactly once. For SQL2S and Parallel, the answer is concise text. For S2SQL, it is the canonical query in a fenced sql block.

These tools are the public interface, not the production retrieval stack used for internal leaderboard runs.