fix: critical bug fixes in query extraction, FAISS search, and MCP config

[zhengdaqi]

Summary

Six bug fixes for correctness issues that produce wrong results or crash at runtime. All fixes are non-breaking for valid inputs.

1. Broken regex in r1_searcher_query_extract (95/100 severity)

The regex pattern r"<|begin_of_query|>([^<]*)" treats | as alternation, so it matches < OR b OR e OR ... instead of the literal token <|begin_of_query|>. This completely breaks query extraction in the Search-R1 pipeline.

Fix: Escape the pipes → r"<\|begin_of_query\|>([^<]*)".

2. Remote MCP server config stored as tuple (90/100)

A trailing comma turns the config into a 1-tuple (dict,) instead of a dict:

# Before (bug): creates a tuple
mcp_servers[name] = ({...},)

# After (fix): creates a dict  
mcp_servers[name] = {...}

This affects both build() and load_pipeline_context(), breaking all HTTP/HTTPS remote MCP server configurations.

3. FAISS search returns wrong passages (85/100)

IndexIDMap.search() returns -1 for padding when fewer than top_k neighbors exist. The code used self.contents[-1], silently returning the last document instead of filtering the padding.

4. _surveycpm_to_one_line() content duplication (78/100)

The function concatenated content with a recursive call on the same content string, producing doubled output text:

# Before: "[OK] content" + "content" (duplicated)
return "[OK] " + string["content"]... + _surveycpm_to_one_line(string["content"])

# After: "[OK] content" (correct)
return "[OK] " + string["content"].replace("\n", " ").strip()

5. Exa web search uses "EMPTY" API key silently (73/100)

When EXA_API_KEY is not configured, the backend silently set api_key="EMPTY", leading to confusing 401 errors at search time. Now raises ToolError with a clear message immediately, consistent with the Tavily backend's behavior.

6. evaluate_trec_pvalue crashes when metrics is None (72/100)

The function signature allows metrics: List[str] | None, but the loop for m in metrics raises TypeError when None is passed. Added a default metrics list.

Test Plan

• Run pytest to verify no regressions
• Verify Search-R1 query extraction with <|begin_of_query|> tokens
• Verify FAISS search with small indices (< top_k documents)
• Test remote MCP server config with HTTP endpoints
• Test Exa backend initialization without API key configured

Made with Cursor