feat: Temporal MCP tool selection for agentic search

[manavgup]

Summary

Enable users to select MCP tools at search time to generate artifacts (PowerPoint, charts, PDFs, etc.) from search results. This uses a temporal/per-request selection model - no database persistence required.

Key Design Decision: Tools are selected per-search request, not configured per-collection. This simplifies implementation and provides maximum flexibility.

Background

This issue implements a simplified approach to agentic RAG that builds on the recently merged MCP Context Forge integration (PR #684). Instead of the database-backed agent configuration described in #697, we start with temporal tool selection and can add persistence later if needed.

Why Temporal First?

Approach	Pros	Cons
Temporal (this issue)	Simple, no DB changes, maximum flexibility	Users re-select tools each time
Persistent (#697)	"Remember my preferences"	Complex, requires DB schema, migrations

We can add optional persistence in a future phase once the core flow is validated.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                      Search Interface                            │
│  ┌───────────────────────────────────────────────────────────┐  │
│  │ 🔍 What were Q4 revenue projections?           [Search]   │  │
│  └───────────────────────────────────────────────────────────┘  │
│                                                                  │
│  Available Tools: (live from MCP Context Forge)                 │
│  ┌────────────┐ ┌────────────┐ ┌────────────┐                  │
│  │ ☑️ PPT Gen │ │ ☑️ Chart   │ │ ☐ PDF      │                  │
│  └────────────┘ └────────────┘ └────────────┘                  │
└──────────────────────────┬──────────────────────────────────────┘
                           │
         POST /api/search  │  { question, collection_id,
                           │    selected_tools: ["ppt-gen", "chart-gen"] }
                           ▼
┌─────────────────────────────────────────────────────────────────┐
│                      SearchService                               │
│  1. Core RAG pipeline → answer + sources                        │
│  2. Execute selected tools (parallel) → artifacts               │
│  3. Return answer + artifacts                                   │
└──────────────────────────┬──────────────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────────────┐
│              MCP Context Forge (already integrated)              │
│              Tools: ppt-generator, chart-generator, etc.        │
└─────────────────────────────────────────────────────────────────┘

Implementation Tasks

Phase 1: Backend (3-4 days)

• Update SearchInput schema to accept selected_tools: list[str]
• Create ToolArtifact and ToolExecution schemas
• Update SearchResponse to include artifacts and tool_executions
• Create ToolExecutorService for parallel tool execution
• Integrate tool execution into SearchService (after answer generation)
• Add GET /api/mcp/tools/available endpoint for tool discovery
• Unit tests for ToolExecutorService (80%+ coverage)
• Integration tests with mock MCP tools

Phase 2: Frontend (2-3 days)

• Create ToolPicker component (fetches available tools, toggle selection)
• Create ArtifactPanel component (display generated artifacts)
• Create ArtifactCard component (preview + download button)
• Integrate ToolPicker into search interface
• Integrate ArtifactPanel into search results
• Handle loading states and errors gracefully
• Responsive design (mobile-friendly)

Phase 3: Sample Tools (2-3 days)

• Create ppt-generator MCP tool for PowerPoint generation
• Create chart-generator MCP tool for data visualization
• Deploy sample tools to MCP Context Forge
• End-to-end testing with real tool invocation
• Documentation for creating custom tools

Dependencies

• PR feat(mcp): Implement MCP Gateway integration for extensibility #684 - MCP Context Forge integration (merged)
• PR feat: SPIFFE/SPIRE Integration Architecture for Agent Identity #695 - SPIFFE/SPIRE agent identity (merged)

Related Issues

• Epic: [EPIC] Future: Agentic AI Platform v2 (Phase 4) #691 - Agentic AI Platform v2
• Supersedes parts of: feat: Implement SearchService 3-stage agent execution hooks #697 (agent hooks - simplified to temporal)
• Defers: feat: Expose RAG Modulo as MCP Server #698 (MCP server exposure - future phase)
• Simplifies: feat: Agentic UI components for agent configuration and artifacts #699 (UI - focused on tool picker + artifacts only)

Acceptance Criteria

• User can see available MCP tools in search interface
• User can select one or more tools before searching
• Search executes normally, then runs selected tools
• Generated artifacts appear in search results
• User can download artifacts (PowerPoint, charts, etc.)
• Tool failures don't break search (graceful degradation)
• Tool execution is parallel (performance)
• Works on desktop and mobile

Future Enhancements (Not in Scope)

These can be added in future issues:

• Optional persistence of tool preferences per collection
• Tool categories in UI
• Tool-specific configuration options
• Pre-search and post-search agent stages (feat: Implement SearchService 3-stage agent execution hooks #697)
• Expose RAG Modulo as MCP server (feat: Expose RAG Modulo as MCP Server #698)

References

• Planning doc: docs/planning/agentic-rag-implementation-plan.md
• MCP integration: docs/features/mcp-integration.md
• Architecture: docs/architecture/search-agent-hooks-architecture.md