feat: Deep Research Multilingual

[yuhongsun96]

Description

Tune it for queries not in English

How Has This Been Tested?

Works for sure

Additional Options

• [Optional] Override Linear Check

---

Summary by cubic

Enable multilingual Deep Research so clarifications, plans, tasks, and reports respond in the user's language. Adds query validation and loop safeguards for more reliable runs.

• New Features

  • Orchestration prompts instruct all steps to respond in the user’s language; research tasks are authored in that language; reports are written in the task language.

• Bug Fixes

  • Normalize web search queries (trim, drop empty) and return a clear error when none are valid.
  • Prevent potential infinite loops by incrementing cycle counters after tool failures; limit tool call logging to debug-only paths.

^{Written for commit b9a1b43. Summary will update on new commits.}

[greptile-apps]

Greptile Summary

Added multilingual support to the deep research feature by instructing the LLM to respond in the user's language at each stage of the research workflow.

• Added "Respond in the user's language" instructions to orchestration layer prompts for clarifications, research plans, research tasks, and final report generation
• Added language matching instructions to research agent prompts to generate reports in the same language as the input task
• Improved web search tool robustness by normalizing queries (stripping whitespace and rejecting empty queries)
• Fixed infinite loop vulnerability by incrementing cycle counters after tool call failures in the research agent
• Minor indentation fix for debug logging

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• Changes are well-targeted prompt engineering additions for multilingual support with solid defensive improvements (query normalization, infinite loop prevention). All changes are additive and non-breaking.
• No files require special attention

Important Files Changed

Filename	Overview
backend/onyx/prompts/deep_research/orchestration_layer.py	Added multilingual support instructions to prompts for clarification, planning, research tasks, and report generation
backend/onyx/prompts/deep_research/research_agent.py	Added language matching instructions to research agent prompts to generate reports in the same language as the input task
backend/onyx/tools/tool_implementations/web_search/web_search_tool.py	Added query normalization to strip whitespace and reject empty queries, improving robustness for multilingual input

Sequence Diagram

sequenceDiagram
    participant User
    participant Orchestrator
    participant ResearchAgent
    participant WebSearch
    participant LLM

    User->>Orchestrator: Submit query (any language)
    Orchestrator->>LLM: Generate clarification questions
    Note over LLM: "Respond in the user's language"
    LLM-->>User: Questions in user's language
    
    User->>Orchestrator: Provide answers
    Orchestrator->>LLM: Generate research plan
    Note over LLM: "Respond in the user's language"
    LLM-->>Orchestrator: Research plan steps (user's language)
    
    loop For each research step
        Orchestrator->>LLM: Create research task
        Note over LLM: "Research task should be in user's language"
        LLM->>ResearchAgent: Research task (user's language)
        
        ResearchAgent->>WebSearch: Execute web searches
        Note over WebSearch: Normalize queries, strip whitespace
        WebSearch-->>ResearchAgent: Search results
        
        ResearchAgent->>LLM: Generate report from results
        Note over LLM: "Write report in same language as task"
        LLM-->>ResearchAgent: Report (user's language)
        ResearchAgent-->>Orchestrator: Research findings
    end
    
    Orchestrator->>LLM: Generate final report
    Note over LLM: "Respond in the user's language"
    LLM-->>User: Final report (user's language)

Loading

[cubic-dev-ai]

No issues found across 5 files

[Subash-Mohan]

lgtm!