Skip to content

implementation_plan.md

Latest commit

 

History

History
56 lines (42 loc) · 5.06 KB

implementation_plan.md

File metadata and controls

56 lines (42 loc) · 5.06 KB
Raw

Implementation Plan

[Overview] The goal is to implement tool-calling capabilities in the OpenInky Word Add-in by natively supporting direct API logic for Fedlex and Onlinekommentar, as well as an MCP client for external servers via SSE.

This implementation enhances the existing LLM integration by introducing a unified ToolService that provides native tools directly translated from the user's Python/Node code and dynamically fetches tools from external MCP servers (like Entscheidsuche). The llmService will be updated to send these tool definitions to the LLM, process tool calls, and feed the results back. This allows legal professionals to directly query Swiss law and commentaries within the Add-in without leaving Microsoft Word. Additionally, licensing and copyright attributions will be added to the project documentation and GUI to acknowledge the incorporated open-source code.

[Types] Introduction of structured types to manage tool execution, LLM requests with tools, and MCP server configurations.

  • ToolDefinition: Interface defining type, function.name, function.description, and function.parameters matching the OpenAI spec.
  • ToolCall: Interface for tool calls received from the LLM, containing id, type, and function.arguments.
  • ToolResponse: Interface for the LLM's required format for submitting tool results (role: 'tool', tool_call_id, content).
  • Updates to LLMConfig in llmService.ts: Add optional mcpServers: string[] to store URLs of external MCP servers.
  • Updates to LLMRequest in llmService.ts: Add optional tools?: ToolDefinition[] and tool_choice?: string.

[Files] Creation of new service files for tool logic and modifications to existing UI and LLM components to support the tool-calling loop.

  • New file src/services/tools/fedlex.ts: Direct translation of the Fedlex Python SPARQL and HTML fetching logic into TypeScript.
  • New file src/services/tools/onlinekommentar.ts: Direct translation of the Onlinekommentar Node API fetching logic into TypeScript.
  • New file src/services/toolService.ts: Central registry for tools, responsible for combining native tools with dynamically fetched MCP tools (via SSEClientTransport) and routing tool executions.
  • Modify src/services/llmService.ts: Update fetchCompletion to accept tools and handle the tool execution loop if tool_calls are returned by the LLM.
  • Modify src/taskpane/components/App.tsx: Update the main prompt execution flow to support tool loading states and pass configurations to the updated llmService.
  • Modify src/taskpane/components/Settings.tsx: Add UI fields for users to configure external MCP server URLs (e.g., https://mcp.entscheidsuche.ch/mcp).
  • Modify README.md and LICENSE.md: Append copyright notices for Fedlex and Onlinekommentar code.

[Functions] Implementation of native tool logic, tool coordination, and recursive LLM completion handling.

  • New functions in fedlex.ts: search_laws, get_law_metadata, get_law_text, list_amendments, browse_taxonomy (making HTTP requests and XML/HTML parsing where necessary).
  • New functions in onlinekommentar.ts: search_commentaries, get_commentary_by_id.
  • New functions in toolService.ts: getAvailableTools(mcpServers: string[]) to aggregate native and remote tools, and executeToolCall(call: ToolCall) to route execution to the correct native or MCP handler.
  • Modified fetchCompletion in llmService.ts: Enhance to detect finish_reason === 'tool_calls', call executeToolCall for each, and recursively call the LLM API with the appended tool responses until completion.

[Classes] Integration of the official Model Context Protocol Client class to manage SSE connections.

  • New class usage in toolService.ts: Instantiate Client from @modelcontextprotocol/sdk/client/index.js and SSEClientTransport from @modelcontextprotocol/sdk/client/sse.js to connect to and interact with configured MCP servers dynamically.

[Dependencies] Installation of required packages for MCP protocol support and parameter validation.

  • Install @modelcontextprotocol/sdk (version ^1.17.1) to enable MCP client connectivity over SSE.
  • Install zod to assist with any schema validation if required by the SDK or tool definitions.

[Implementation Order] Sequential integration starting from core API logic, building up to LLM integration and UI configuration.

  1. Install new dependencies (@modelcontextprotocol/sdk and zod).
  2. Create src/services/tools/fedlex.ts and src/services/tools/onlinekommentar.ts with direct API fetching logic.
  3. Create src/services/toolService.ts to manage native tools and establish the MCP Client logic for external servers.
  4. Update src/services/llmService.ts to support OpenAI tool schemas, parsing tool_calls, executing them via toolService, and handling the recursive tool-call completion loop.
  5. Update src/taskpane/components/Settings.tsx to add UI for entering MCP Server URLs and display copyright/license notices.
  6. Update README.md and LICENSE.md with appropriate attributions.
  7. Test the full tool-calling flow with App.tsx handling complex legal queries.