Skip to main content

Deep Analysis Improvements

This document describes the improvements made to Radium's AI analysis capabilities to ensure agents perform comprehensive, multi-file analysis instead of shallow, single-file responses.

Problem​

Previously, when asked general questions like "Tell me about this project", agents would:

  • Read only one file (often just GEMINI.md or a single doc file)
  • Give surface-level answers without deep understanding
  • Skip comprehensive analysis
  • Not synthesize information from multiple sources

Solution​

1. Enhanced Agent Prompts​

All key agents now have mandatory deep analysis protocols:

  • Research Agent: 5-phase analysis protocol with mandatory file reading
  • Code Agent: Pre-implementation analysis workflow
  • Analyzer Agent: Comprehensive analysis with multi-tool coordination

Key improvements:

  • MANDATORY language - agents are explicitly told they MUST follow protocols
  • Parallel file reading instructions
  • Introspection checklists that must be completed before answering
  • Explicit prohibition of single-file answers

2. Question-Type Detection​

Created QuestionType enum and AnalysisPlan system that:

  • Detects question types (ProjectOverview, TechnologyStack, Architecture, etc.)
  • Recommends specific files to read for each question type
  • Suggests semantic search queries
  • Provides synthesis guidance

3. Analysis Plan Integration​

Enhanced execute_chat_message in TUI to:

  • Automatically create analysis plans for user questions
  • Inject analysis plans into agent prompts
  • Prepend analysis guidance to prompt content

4. Context Manager Enhancements​

Added to ContextManager:

  • create_analysis_plan() - Creates analysis plans from user input
  • build_context_with_analysis() - Builds context with analysis plan included

Usage​

Automatic (TUI Chat)​

When using TUI chat (/chat research-agent or orchestration), analysis plans are automatically:

  1. Created from user input
  2. Injected into the prompt
  3. Enforced by agent instructions

Manual (CLI)​

You can use analysis plans programmatically:

use radium_core::context::{ContextManager, Workspace};

let workspace = Workspace::discover()?;
let manager = ContextManager::new(&workspace);
let plan = manager.create_analysis_plan("Tell me about this project");

// Use plan.recommended_files, plan.suggested_searches, etc.

Expected Behavior​

When asked "Tell me about this project", agents should now:

  1. Read multiple files in parallel:

    • README.md
    • package.json / Cargo.toml
    • nx.json / rust-toolchain.toml
    • Architecture docs
    • GEMINI.md

  2. Perform semantic searches:

    • "What is the main purpose and architecture of this project?"
    • Related architecture queries

  3. Synthesize information:

    • Combine findings from all sources
    • Provide comprehensive overview
    • Include specific examples with file paths

  4. Verify completeness:

    • Complete introspection checklist
    • Ensure all aspects covered
    • Show deep understanding

Files Modified​

  • prompts/agents/specialized/research-agent.md - Enhanced with mandatory protocols
  • prompts/agents/core/code-agent.md - Added analysis workflows
  • prompts/agents/specialized/analyzer-agent.md - Added introspection
  • apps/tui/src/chat_executor.rs - Integrated analysis plans
  • crates/radium-core/src/context/analysis.rs - Question type detection
  • crates/radium-core/src/context/manager.rs - Analysis plan methods
  • crates/radium-orchestrator/src/routing/question_type.rs - Orchestrator question types

Testing​

To verify the improvements work:

  1. Ask a general question: "Tell me about this project"
  2. Check that the agent reads multiple files (check tool calls)
  3. Verify the answer is comprehensive and includes:
    • Technology stack details
    • Architecture information
    • Specific file references
    • Multiple sources synthesized

Future Enhancements​

  • Pre-execution file reading phase (force file reads before agent execution)
  • Agent routing based on question type (route general questions to research-agent)
  • Analysis plan caching for similar questions
  • Metrics tracking for analysis depth