AI MCP Gateway

/** * @file Orchestrator Schemas * @description Zod schemas for structured LLM responses in multi-pass orchestration */ import { z } from 'zod'; /** * Schema for analysis pass output * LLM extracts intent, queries, and determines if retrieval is needed */ export const AnalysisResponseSchema = z.object({ intent: z.string().describe('The user intent or main goal of the request'), keywords: z.array(z.string()).describe('Key search terms for context retrieval'), requires_retrieval: z.boolean().describe('Whether external context/history is needed'), requires_tools: z.boolean().optional().describe('Whether tool calls are needed'), complexity_level: z.enum(['simple', 'moderate', 'complex']).describe('Estimated complexity'), approach: z.string().describe('Suggested approach to solve this request'), clarifications: z.array(z.string()).optional().describe('Questions needing clarification'), }); export type AnalysisResponse = z.infer<typeof AnalysisResponseSchema>; /** * Schema for generation pass output * LLM generates the main response with citations */ export const GenerationResponseSchema = z.object({ response: z.string().describe('Main response to the user query'), citations: z.array(z.object({ reference: z.string().describe('Source or evidence reference'), text: z.string().describe('Cited text'), })).optional().describe('Citations from provided context'), confidence: z.number().min(0).max(1).optional().describe('Confidence level of the response'), requires_follow_up: z.boolean().optional().describe('Whether follow-up questions should be asked'), }); export type GenerationResponse = z.infer<typeof GenerationResponseSchema>; /** * Schema for refinement pass output * LLM improves draft response based on criteria */ export const RefinementResponseSchema = z.object({ refined_response: z.string().describe('Refined and improved response'), improvements_made: z.array(z.string()).describe('List of improvements applied'), quality_score: z.number().min(0).max(100).optional().describe('Quality score of refined response'), ready_to_send: z.boolean().describe('Whether response is ready to send to user'), }); export type RefinementResponse = z.infer<typeof RefinementResponseSchema>; /** * Configuration for orchestrator strategy */ export const OrchestratorStrategySchema = z.object({ enabled: z.boolean().default(false).describe('Enable multi-pass orchestration'), passes: z.enum(['two-pass', 'three-pass']).default('two-pass').describe('Number of LLM passes'), include_analysis: z.boolean().default(true).describe('Include analysis pass'), include_refinement: z.boolean().default(false).describe('Include refinement pass'), max_total_tokens: z.number().default(8000).describe('Max tokens for entire orchestration'), timeout_ms: z.number().default(30000).describe('Timeout for orchestration'), }); export type OrchestratorStrategy = z.infer<typeof OrchestratorStrategySchema>; /** * Orchestration result with all passes' outputs */ export const OrchestratorResultSchema = z.object({ final_response: z.string().describe('Final response to send to user'), analysis: AnalysisResponseSchema.optional(), draft: GenerationResponseSchema.optional(), refined: RefinementResponseSchema.optional(), token_usage: z.object({ analysis_tokens: z.number().default(0), generation_tokens: z.number().default(0), refinement_tokens: z.number().default(0), total_tokens: z.number(), }), duration_ms: z.number(), passes_completed: z.array(z.string()), }); export type OrchestratorResult = z.infer<typeof OrchestratorResultSchema>;