Sonar Reasoning Pro (reason tool) returns irrelevant search results

[SangeethsivanSivakumar]

Bug Description

The reason tool (Sonar Reasoning Pro) consistently returns irrelevant search results, while the search tool (Sonar Pro) works correctly for the same queries. The reasoning model appears to receive garbage search results from its backend, causing it to either answer from training data (ignoring sources) or correctly refuse to answer.

Reproduction

Same query, both tools, side by side:

search (works correctly)

Query: Cleveland Fed inflation nowcast March 2026 CPI estimate
Result: Correct answer (0.84% MoM, cites clevelandfed.org)
Sources: Relevant (clevelandfed.org, econbrowser.com, fred.stlouisfed.org)

reason (broken)

Query: Cleveland Fed inflation nowcast March 2026 CPI estimate
Result: "I cannot answer this query based on the search results provided"
Sources: Irrelevant (axios.com/local/cleveland, news5cleveland.com, clevelandohio.gov - local Cleveland news, not Federal Reserve data)

Pattern

Tested 5+ queries across a session. reason consistently returns sources from a narrow set of irrelevant domains:

Query	reason Sources Returned
Capital of France	situational interview questions, relevance examples
Cleveland Fed nowcast	local Cleveland city news
March 2026 CPI forecast	interview questions, error handling tutorials
Seoul weather forecast	interview questions, workplace scenarios
Trump Truth Social posting	error handling strategies, debugging tutorials

The model is smart enough to answer from training data when it can (e.g., "Paris"), but fails on queries requiring search results (e.g., real-time economic data).

Environment

• Client: Claude Code (Anthropic CLI)
• MCP Server: Perplexity MCP (via Claude Code plugin)
• Date: April 7, 2026
• search tool: Works perfectly for all the same queries
• reason tool: Broken across all tested queries

Expected Behavior

reason should receive the same quality of search results as search, then apply deeper reasoning on top.

Actual Behavior

reason receives unrelated search results (appears to be a different/broken search query being sent to the backend), causing the reasoning model to either hallucinate an answer from training data or correctly refuse.

Workaround

Use search (Sonar Pro) for all queries. It handles both simple lookups and synthesis/reasoning questions effectively.

[SangeethsivanSivakumar]

Root Cause Update

Debugged the MCP wrapper (perplexity-mcp npm package v0.2.3 from DaInfernalCoder/researcher-mcp). The wrapper code is NOT the issue. Both search and reason use identical API call patterns:

const response = await this.axiosInstance.post("/chat/completions", {
    model,  // "sonar-pro" for search, "sonar-reasoning-pro" for reason
    messages: [{ role: "user", content: prompt }],
});

Same endpoint, same message structure, only the model name differs. sonar-pro returns relevant search results; sonar-reasoning-pro returns garbage for the same queries.

However, testing in Cursor IDE with a different MCP server that exposes search_context_size and search_domain_filter parameters (which the npm package does not pass through), sonar-reasoning-pro works correctly.

Likely fix: The sonar-reasoning-pro model may require search_context_size to be explicitly set (e.g., "high") to produce relevant search results via the API, while sonar-pro defaults to sensible search behavior without it. The npm wrapper doesn't expose or pass these parameters.

This could be fixed either:

1. In the API: Make sonar-reasoning-pro default to search_context_size: "high" when not specified
2. In the npm wrapper: Pass through search_context_size and other search parameters to the API call

[SangeethsivanSivakumar]

Confirmed Root Cause: Prompt Template Poisons Search

Systematically debugged with direct curl calls to the Perplexity API.

The bug is in the prompt template wrapping, NOT in search_context_size.

The reason tool wraps every query in this template before sending to the API:

You are answering a query that contains specific details like error messages, 
logs, code snippets, exact terminology, version numbers, and context. 
Carefully analyze all provided details...

Query: {actual_user_query}

Provide a detailed explanation and analysis that:
1. Addresses the specific details provided (errors, logs, code, versions, etc.)
...

Perplexity's search component extracts search terms from the FULL message content. The template keywords ("error messages", "logs", "code snippets", "specific situation", "context") dominate the search query, returning results about interview questions and problem-solving instead of the user's actual topic.

Test matrix (all using sonar-reasoning-pro via direct API curl):

Prompt	search_context_size	Citations
Clean query only	none	✅ Relevant (clevelandfed.org)
Clean query only	high	✅ Relevant (clevelandfed.org)
Wrapper template + query	none	❌ Garbage (interview sites)
Wrapper template + query	high	❌ Garbage (interview sites)

sonar-pro (used by search) is more robust to prompt noise, which is why search works despite using the same template.

Fix: Strip the template and pass the raw user query directly. The models work perfectly without prompt wrapping.

case "reason": {
    model = "sonar-reasoning-pro";
-   prompt = `You are answering a query that contains specific details...
-   Query: ${query}
-   Provide a detailed explanation...`;
+   prompt = query;
    break;
}

This should also be applied to the search tool as a preventive measure.

[SangeethsivanSivakumar]

Fix Confirmed

Patched locally and verified end-to-end through the MCP tool in Claude Code.

Before patch (via MCP tool)

Query: "What is the Cleveland Fed inflation nowcast for March 2026 CPI?"
Citations: skillstracking.com, noota.io, indeed.com (interview questions)
Answer: "I cannot answer this query based on the search results provided"

After patch (via MCP tool)

Query: Same
Citations: clevelandfed.org, econbrowser.com, fred.stlouisfed.org
Answer: Correct (0.84% MoM, energy-driven, cites nowcasting methodology)

Also tested with a complex reasoning query ("strongest arguments March 2026 CPI could exceed 3.4%") - returned relevant sources (octagonai.co, tradingeconomics.com, bls.gov, federalreserve.gov) and substantive analysis.

The fix (2 lines changed)

// reason tool (line ~306)
case "reason": {
    model = "sonar-reasoning-pro";
-   prompt = `You are answering a query that contains specific details like error messages, logs, code snippets, exact terminology, version numbers, and context. Carefully analyze all provided details to give the most accurate and helpful answer.\n\nQuery: ${query}\n\nProvide a detailed explanation and analysis that:\n1. Addresses the specific details provided (errors, logs, code, versions, etc.)\n2. Includes step-by-step reasoning based on the actual context\n3. Identifies key considerations relevant to the specific situation\n4. Provides relevant examples matching the described scenario\n5. Offers practical implications based on the exact details provided\n6. Suggests potential alternatives or solutions tailored to the specific context`;
+   prompt = query;
    break;
}

// search tool (line ~297) - preventive, sonar-pro is more robust but same issue applies
case "search": {
    model = "sonar-pro";
-   prompt = `You are answering a query that contains specific details like error messages, logs, code snippets, exact terminology, version numbers, and context. Use all provided details to give the most accurate answer possible.\n\nQuery: ${query}\n\nProvide a clear, concise answer that directly addresses the specific details in the query.`;
+   prompt = query;
    break;
}

The models handle raw queries perfectly without prompt wrapping. The template was actively harmful because Perplexity's search component extracts search terms from the full message content, and the template keywords dominated over the actual query.