@framers/agentos 0.6.30 → 0.7.1

package/dist/orchestration/pipeline/query/QueryRouter.js

@@ -0,0 +1,1762 @@
+/**
+ * @fileoverview QueryRouter — main orchestrator that wires together the
+ * QueryClassifier, QueryDispatcher, and QueryGenerator into a complete
+ * classify -> dispatch -> generate pipeline.
+ *
+ * @module @framers/agentos/query-router/QueryRouter
+ *
+ * The QueryRouter is the top-level entry point for the intelligent query
+ * routing system. Given a user query and optional conversation history, it:
+ *
+ * 1. **Classifies** the query into a complexity tier (T0-T3) using the
+ *    {@link QueryClassifier}.
+ * 2. **Dispatches** the classified query to the tier-appropriate retrieval
+ *    pipeline via the {@link QueryDispatcher}, which orchestrates vector
+ *    search, graph expansion, reranking, and deep research.
+ * 3. **Generates** a grounded answer with citations via the
+ *    {@link QueryGenerator}, using the retrieved context chunks.
+ *
+ * The router also handles:
+ * - Corpus loading from markdown files on disk
+ * - Real vector embedding via EmbeddingManager + VectorStoreManager (in-memory)
+ * - Keyword fallback retrieval when embeddings are unavailable
+ * - Event emission for full pipeline observability
+ * - Lifecycle hooks (onClassification, onRetrieval) for consumer integration
+ *
+ * Embedding pipeline: The `init()` method attempts to initialize an
+ * AIModelProviderManager, EmbeddingManager, and VectorStoreManager to enable
+ * real vector search. If any step fails (e.g., no API key configured), the
+ * router falls back gracefully to {@link KeywordFallback} for all retrieval.
+ */
+import { existsSync, readdirSync, readFileSync } from 'node:fs';
+import { dirname, extname, join, resolve } from 'node:path';
+import { fileURLToPath, pathToFileURL } from 'node:url';
+import { QueryClassifier } from './QueryClassifier.js';
+import { QueryDispatcher } from './QueryDispatcher.js';
+import { QueryGenerator } from './QueryGenerator.js';
+import { TopicExtractor } from './TopicExtractor.js';
+import { KeywordFallback } from './KeywordFallback.js';
+import { CitationVerifier } from '../../../cognition/rag/citation/CitationVerifier.js';
+import { DEFAULT_QUERY_ROUTER_CONFIG, DEFAULT_STRATEGY_CONFIG, } from './types.js';
+/** Regex for splitting markdown by h1-h3 headings. */
+const HEADING_REGEX = /^#{1,3}\s+(.+)/;
+/** Maximum character length for a single corpus chunk. */
+const MAX_CHUNK_CHARS = 6000;
+/** Minimum content length for a chunk to be included. */
+const MIN_CHUNK_CHARS = 20;
+/** Supported markdown file extensions for corpus loading. */
+const MARKDOWN_EXTENSIONS = new Set(['.md', '.mdx']);
+const MODULE_DIR = dirname(fileURLToPath(import.meta.url));
+const GITHUB_EXTENSION_LOCAL_ENTRY_CANDIDATES = [
+    resolve(MODULE_DIR, '../../../../packages/agentos-extensions/registry/curated/integrations/github/dist/index.js'),
+    resolve(MODULE_DIR, '../../../../apps/wunderland-sol/packages/agentos-extensions/registry/curated/integrations/github/dist/index.js'),
+];
+// ============================================================================
+// QueryRouter
+// ============================================================================
+/**
+ * Main orchestrator that wires together the QueryClassifier, QueryDispatcher,
+ * and QueryGenerator into a complete classify -> dispatch -> generate pipeline.
+ *
+ * @example
+ * ```typescript
+ * const router = new QueryRouter({
+ *   knowledgeCorpus: ['./docs'],
+ *   generationModel: 'gpt-4o-mini',
+ *   generationModelDeep: 'gpt-4o',
+ *   generationProvider: 'openai',
+ * });
+ *
+ * await router.init();
+ * const result = await router.route('How does authentication work?');
+ * console.log(result.answer);
+ * console.log(result.sources);
+ *
+ * await router.close();
+ * ```
+ */
+export class QueryRouter {
+    /**
+     * Creates a new QueryRouter instance.
+     *
+     * Merges user-supplied configuration over {@link QUERY_ROUTER_DEFAULTS}.
+     * The router is NOT ready to use until {@link init} is called.
+     *
+     * @param config - Partial configuration; `knowledgeCorpus` is required.
+     */
+    constructor(config) {
+        /** Loaded corpus chunks from disk. */
+        this.corpus = [];
+        /** Topic entries extracted from the corpus. */
+        this.topics = [];
+        /** Keyword-based fallback search engine. */
+        this.keywordFallback = null;
+        /** Chain-of-thought query classifier. */
+        this.classifier = null;
+        /**
+         * Optional capability discovery engine persisted across classifier rebuilds.
+         *
+         * The router recreates its classifier during `init()` and when background
+         * GitHub indexing refreshes the topic list. Persist the attached discovery
+         * engine so those classifier rebuilds do not silently drop capability-aware
+         * planning.
+         */
+        this.capabilityDiscoveryEngine = null;
+        /** Tier-routing dispatcher. */
+        this.dispatcher = null;
+        /** LLM answer generator. */
+        this.generator = null;
+        /** Accumulated lifecycle events for observability. */
+        this.events = [];
+        /** In-memory cache for completed route() results. */
+        this.routeResultCache = new Map();
+        /** Whether init() has been called successfully. */
+        this.initialized = false;
+        /** Embedding manager for generating vector embeddings. Null if not available. */
+        this.embeddingManager = null;
+        /** Vector store manager for persisting and querying embeddings. Null if not available. */
+        this.vectorStoreManager = null;
+        /** AI model provider manager used by the embedding manager. Null if not available. */
+        this.providerManager = null;
+        /** Embedding dimension for the configured model. Zero if embeddings unavailable. */
+        this.embeddingDimension = 0;
+        /** Current embedding availability state for corpus retrieval. */
+        this.embeddingStatus = 'disabled-no-key';
+        /**
+         * Optional UnifiedRetriever for plan-based retrieval.
+         *
+         * When set via {@link setUnifiedRetriever}, the `route()` method uses
+         * the UnifiedRetriever instead of the legacy QueryDispatcher for the
+         * retrieval phase. The UnifiedRetriever executes a structured
+         * {@link RetrievalPlan} across all available sources in parallel.
+         *
+         * @see setUnifiedRetriever
+         */
+        this.unifiedRetriever = null;
+        /**
+         * The data source ID used for corpus embeddings in the vector store.
+         * Matches the collection name configured during init().
+         */
+        this.corpusDataSourceId = 'query-router-corpus';
+        this.config = {
+            ...DEFAULT_QUERY_ROUTER_CONFIG,
+            ...config,
+            deepResearchEnabled: config.deepResearchEnabled ?? Boolean(process.env.SERPER_API_KEY),
+            availableTools: config.availableTools ?? [...DEFAULT_QUERY_ROUTER_CONFIG.availableTools],
+            strategyConfig: {
+                ...DEFAULT_STRATEGY_CONFIG,
+                ...config.strategyConfig,
+            },
+        };
+    }
+    // ==========================================================================
+    // UNIFIED RETRIEVER INTEGRATION
+    // ==========================================================================
+    /**
+     * Attach a {@link UnifiedRetriever} for plan-based retrieval.
+     *
+     * When set, the `route()` method uses the UnifiedRetriever instead of
+     * the legacy QueryDispatcher for the retrieval phase. The classifier
+     * automatically produces a {@link RetrievalPlan} via `classifyWithPlan()`
+     * and the retriever executes it across all available sources in parallel.
+     *
+     * Pass `null` to revert to the legacy QueryDispatcher pipeline.
+     *
+     * @param retriever - A configured UnifiedRetriever instance, or `null` to disable.
+     *
+     * @example
+     * ```typescript
+     * const retriever = new UnifiedRetriever({
+     *   hybridSearcher, raptorTree, graphEngine, memoryManager,
+     * });
+     * router.setUnifiedRetriever(retriever);
+     * // Now route() uses plan-based retrieval automatically
+     * ```
+     */
+    setUnifiedRetriever(retriever) {
+        this.unifiedRetriever = retriever;
+        this.clearRouteResultCache();
+    }
+    /**
+     * Get the attached UnifiedRetriever, or `null` if not configured.
+     *
+     * @returns The UnifiedRetriever instance, or `null`.
+     */
+    getUnifiedRetriever() {
+        return this.unifiedRetriever;
+    }
+    // ==========================================================================
+    // CAPABILITY DISCOVERY INTEGRATION
+    // ==========================================================================
+    /**
+     * Attach a {@link CapabilityDiscoveryEngine} for capability-aware classification.
+     *
+     * When set, the classifier injects Tier 0 capability summaries (~150 tokens)
+     * into its LLM prompt, enabling it to recommend which skills, tools, and
+     * extensions should be activated for each query. The recommendations are
+     * included in the {@link ExecutionPlan} returned by `classifyWithPlan()`.
+     *
+     * Pass `null` to detach and revert to keyword-based heuristic capability
+     * selection.
+     *
+     * @param engine - A configured and initialized CapabilityDiscoveryEngine, or `null` to detach.
+     *
+     * @example
+     * ```typescript
+     * const engine = new CapabilityDiscoveryEngine(embeddingManager, vectorStore);
+     * await engine.initialize({ tools, skills, extensions, channels });
+     * router.setCapabilityDiscoveryEngine(engine);
+     * // Now route() includes skill/tool/extension recommendations in the execution plan
+     * ```
+     */
+    setCapabilityDiscoveryEngine(engine) {
+        this.capabilityDiscoveryEngine = engine;
+        if (this.classifier) {
+            this.classifier.setCapabilityDiscoveryEngine(engine);
+        }
+        this.clearRouteResultCache();
+    }
+    // ==========================================================================
+    // PUBLIC API
+    // ==========================================================================
+    /**
+     * Initialise the router: load corpus from disk, extract topics, build
+     * keyword fallback index, embed the corpus into a vector store, and
+     * instantiate classifier/dispatcher/generator.
+     *
+     * Must be called before `classify()`, `retrieve()`, or `route()`.
+     *
+     * The embedding step uses real EmbeddingManager + VectorStoreManager when
+     * an LLM provider is available (e.g., OPENAI_API_KEY is set). If embedding
+     * initialisation fails for any reason, the router falls back gracefully to
+     * KeywordFallback for all retrieval.
+     */
+    async init() {
+        // 1. Load corpus chunks from the configured knowledge directories
+        this.corpus = this.loadCorpus(this.config.knowledgeCorpus);
+        // 1b. Load bundled platform knowledge (tools, skills, FAQ, API, troubleshooting)
+        if (this.config.includePlatformKnowledge !== false) {
+            const platformChunks = this.loadPlatformKnowledge();
+            if (platformChunks.length > 0) {
+                this.corpus.push(...platformChunks);
+            }
+        }
+        if (this.corpus.length === 0) {
+            throw new Error(this.buildEmptyCorpusError(this.config.knowledgeCorpus));
+        }
+        // 2. Extract topics for the classifier's system prompt
+        const topicExtractor = new TopicExtractor();
+        this.topics = topicExtractor.extract(this.corpus);
+        const topicList = topicExtractor.formatForPrompt(this.topics);
+        // 3. Build keyword fallback index
+        this.keywordFallback = new KeywordFallback(this.corpus);
+        // 4. Attempt to embed corpus chunks into a real vector store.
+        //    This is wrapped in a try/catch so failure is non-fatal — keyword
+        //    fallback will still work for all retrieval operations.
+        await this.embedCorpus();
+        // 5. Instantiate the classifier
+        this.classifier = this.createClassifier(topicList);
+        // 6. Instantiate the generator
+        this.generator = new QueryGenerator({
+            model: this.config.generationModel,
+            modelDeep: this.config.generationModelDeep,
+            provider: this.config.generationProvider,
+            apiKey: this.getLlmApiKey(),
+            baseUrl: this.getLlmBaseUrl(),
+            maxContextTokens: this.config.maxContextTokens,
+        });
+        // 7. Instantiate the dispatcher with callback dependencies
+        this.dispatcher = new QueryDispatcher({
+            vectorSearch: (query, topK) => this.vectorSearch(query, topK),
+            hydeSearch: (query, topK) => this.hydeSearch(query, topK),
+            decompose: (query, maxSubQueries) => this.decomposeQuery(query, maxSubQueries),
+            graphExpand: (seeds) => this.config.graphExpand ? this.config.graphExpand(seeds) : this.graphExpand(seeds),
+            rerank: (query, chunks, topN) => this.config.rerank
+                ? this.config.rerank(query, chunks, topN)
+                : this.rerank(query, chunks, topN),
+            deepResearch: (query, sources) => this.config.deepResearch
+                ? this.config.deepResearch(query, sources)
+                : this.deepResearch(query, sources),
+            emit: (event) => this.emit(event),
+            graphEnabled: this.config.graphEnabled,
+            deepResearchEnabled: this.config.deepResearchEnabled,
+            maxSubQueries: this.config.strategyConfig.maxSubQueries,
+        });
+        this.initialized = true;
+        // ----------------------------------------------------------------
+        // Background GitHub repo indexing (non-blocking)
+        // ----------------------------------------------------------------
+        // Runs after the router is fully initialized so that query routing
+        // is available immediately. Indexed chunks are merged into the
+        // corpus, then the live retrieval/classification state is refreshed
+        // once indexing completes.
+        const repoConfig = this.config.githubRepos ?? {};
+        const includeEcosystem = repoConfig.includeEcosystem ?? true;
+        if (includeEcosystem || (repoConfig.repos && repoConfig.repos.length > 0)) {
+            Promise.resolve().then(async () => {
+                try {
+                    const { GitHubRepoIndexer, GitHubService } = await this.loadGitHubExtensionModule();
+                    const indexedChunks = [];
+                    const token = repoConfig.token || process.env.GITHUB_TOKEN || process.env.GH_TOKEN || '';
+                    const service = new GitHubService(token);
+                    // initialize() creates the Octokit and validates auth.
+                    // If the token is empty/invalid, Octokit is still created before
+                    // the auth check throws, so public-only API access still works.
+                    try {
+                        await service.initialize();
+                    }
+                    catch { /* public-only mode */ }
+                    const indexer = new GitHubRepoIndexer(service);
+                    if (includeEcosystem) {
+                        const results = await indexer.indexEcosystem();
+                        const totalChunks = results.reduce((s, r) => s + r.chunks.length, 0);
+                        for (const result of results) {
+                            indexedChunks.push(...result.chunks);
+                            this.emit({
+                                type: 'github:index:complete',
+                                repo: result.repo,
+                                chunksTotal: result.chunks.length,
+                                durationMs: result.durationMs,
+                                timestamp: Date.now(),
+                            });
+                        }
+                        console.log(`[QueryRouter] GitHub ecosystem indexed: ${results.length} repos, ${totalChunks} chunks`);
+                    }
+                    if (repoConfig.repos?.length) {
+                        for (const { owner, repo } of repoConfig.repos) {
+                            try {
+                                const result = await indexer.indexRepo(owner, repo);
+                                indexedChunks.push(...result.chunks);
+                                this.emit({
+                                    type: 'github:index:complete',
+                                    repo: result.repo,
+                                    chunksTotal: result.chunks.length,
+                                    durationMs: result.durationMs,
+                                    timestamp: Date.now(),
+                                });
+                            }
+                            catch (err) {
+                                this.emit({
+                                    type: 'github:index:error',
+                                    repo: `${owner}/${repo}`,
+                                    error: err instanceof Error ? err.message : String(err),
+                                    timestamp: Date.now(),
+                                });
+                            }
+                        }
+                    }
+                    await this.syncIndexedCorpusChunks(indexedChunks);
+                }
+                catch (err) {
+                    console.warn(`[QueryRouter] GitHub indexing failed: ${err instanceof Error ? err.message : String(err)}`);
+                }
+            });
+        }
+    }
+    async loadGitHubExtensionModule() {
+        const specifiers = [
+            '@framers/agentos-ext-github',
+            ...GITHUB_EXTENSION_LOCAL_ENTRY_CANDIDATES
+                .filter((candidate) => existsSync(candidate))
+                .map((candidate) => pathToFileURL(candidate).href),
+        ];
+        const failures = [];
+        for (const specifier of specifiers) {
+            try {
+                const module = (await import(/* @vite-ignore */ specifier));
+                if (module.GitHubRepoIndexer && module.GitHubService) {
+                    return module;
+                }
+                failures.push(`${specifier} loaded without GitHubRepoIndexer/GitHubService exports`);
+            }
+            catch (error) {
+                failures.push(`${specifier}: ${error instanceof Error ? error.message : String(error)}`);
+            }
+        }
+        throw new Error(`Unable to load GitHub extension runtime for QueryRouter indexing. ${failures.join(' | ')}`);
+    }
+    /**
+     * Classify a query into a complexity tier without dispatching or generating.
+     *
+     * Useful when consumers want to inspect the classification before deciding
+     * whether to proceed with the full pipeline.
+     *
+     * @param query - The user's natural-language query.
+     * @param conversationHistory - Optional recent conversation messages.
+     * @returns The classification result with tier, confidence, and reasoning.
+     * @throws If the router has not been initialised via {@link init}.
+     */
+    async classify(query, conversationHistory, options) {
+        this.ensureInitialized();
+        const start = Date.now();
+        this.emit({
+            type: 'classify:start',
+            query,
+            timestamp: start,
+        });
+        const trimmedHistory = this.trimConversationHistory(conversationHistory);
+        const result = await this.classifier.classify(query, trimmedHistory, options);
+        if (result.reasoning.startsWith('Classification failed;')) {
+            this.emit({
+                type: 'classify:error',
+                error: new Error(result.reasoning),
+                timestamp: Date.now(),
+            });
+        }
+        this.emit({
+            type: 'classify:complete',
+            result,
+            durationMs: Date.now() - start,
+            timestamp: Date.now(),
+        });
+        return result;
+    }
+    /**
+     * Retrieve context at a specific tier, bypassing the classifier.
+     *
+     * Useful when the caller already knows the appropriate retrieval depth
+     * and wants to skip classification overhead.
+     *
+     * @param query - The user's natural-language query.
+     * @param tier - The complexity tier to retrieve at (0-3).
+     * @returns The retrieval result with chunks and optional graph/research data.
+     * @throws If the router has not been initialised via {@link init}.
+     */
+    async retrieve(query, tier) {
+        this.ensureInitialized();
+        return this.dispatcher.dispatch(query, tier);
+    }
+    /**
+     * Full end-to-end pipeline: classify -> dispatch -> generate.
+     *
+     * This is the primary method for answering user queries. It:
+     * 1. Classifies the query to determine retrieval depth.
+     * 2. Dispatches retrieval at the classified tier.
+     * 3. Generates a grounded answer from the retrieved context.
+     * 4. Emits lifecycle events throughout for observability.
+     *
+     * @param query - The user's natural-language query.
+     * @param conversationHistory - Optional recent conversation messages.
+     * @returns The final query result with answer, classification, sources, and timing.
+     * @throws If the router has not been initialised via {@link init}.
+     */
+    async route(query, conversationHistory, options) {
+        this.ensureInitialized();
+        const routeStart = Date.now();
+        const trimmedHistory = this.trimConversationHistory(conversationHistory);
+        const cacheKey = this.buildRouteCacheKey(query, trimmedHistory, options);
+        if (this.config.cacheResults !== false) {
+            const cached = this.routeResultCache.get(cacheKey);
+            if (cached) {
+                const cachedResult = this.cloneQueryResult(cached);
+                this.emit({
+                    type: 'route:complete',
+                    result: cachedResult,
+                    durationMs: Date.now() - routeStart,
+                    timestamp: Date.now(),
+                });
+                return cachedResult;
+            }
+        }
+        // --- Phase 1: Classification ---
+        const classifyStart = Date.now();
+        this.emit({
+            type: 'classify:start',
+            query,
+            timestamp: classifyStart,
+        });
+        const [classification, executionPlan] = await this.classifier.classifyWithPlan(query, trimmedHistory, options);
+        if (classification.reasoning.startsWith('Classification failed;')) {
+            this.emit({
+                type: 'classify:error',
+                error: new Error(classification.reasoning),
+                timestamp: Date.now(),
+            });
+        }
+        this.emit({
+            type: 'classify:complete',
+            result: classification,
+            durationMs: Date.now() - classifyStart,
+            timestamp: Date.now(),
+        });
+        // Fire the onClassification hook if configured
+        if (this.config.onClassification) {
+            this.config.onClassification(classification);
+        }
+        // The router recommends capabilities but does not activate them. Emit the
+        // suggestion event for both legacy and unified retrieval paths so hosts can
+        // honor recommendations consistently.
+        if (executionPlan.skills.length > 0 ||
+            executionPlan.tools.length > 0 ||
+            executionPlan.extensions.length > 0) {
+            this.emit({
+                type: 'capabilities:activate',
+                skills: executionPlan.skills,
+                tools: executionPlan.tools,
+                extensions: executionPlan.extensions,
+                timestamp: Date.now(),
+            });
+        }
+        // --- Phase 2: Retrieval (with capability recommendations) ---
+        // When a UnifiedRetriever is attached, use plan-based retrieval with
+        // full ExecutionPlan (including skill/tool/extension recommendations).
+        // Otherwise fall back to the legacy QueryDispatcher pipeline.
+        const retrievalEventStart = this.events.length;
+        let retrieval;
+        if (this.unifiedRetriever && this.classifier) {
+            // Plan-based retrieval via UnifiedRetriever
+            const unifiedResult = await this.unifiedRetriever.retrieve(query, executionPlan);
+            retrieval = {
+                chunks: unifiedResult.chunks,
+                researchSynthesis: unifiedResult.researchSynthesis,
+                durationMs: unifiedResult.durationMs,
+            };
+        }
+        else {
+            // Legacy dispatcher pipeline (HyDE-aware strategy or tier-based)
+            retrieval = classification.strategy
+                ? await this.dispatcher.dispatchByStrategy(query, classification.strategy, classification.suggestedSources)
+                : await this.dispatcher.dispatch(query, classification.tier, classification.suggestedSources);
+        }
+        const retrievalEvents = this.events.slice(retrievalEventStart);
+        const fallbacksUsed = this.collectFallbacks(classification, retrievalEvents);
+        const tiersUsed = this.collectTiersUsed(classification, fallbacksUsed);
+        // Fire the onRetrieval hook if configured
+        if (this.config.onRetrieval) {
+            this.config.onRetrieval(retrieval);
+        }
+        // --- Phase 3: Generation ---
+        this.emit({
+            type: 'generate:start',
+            contextChunkCount: retrieval.chunks.length,
+            timestamp: Date.now(),
+        });
+        const generateResult = await this.generator.generate(query, classification.tier, retrieval.chunks, retrieval.researchSynthesis);
+        // Build source citations from the retrieved chunks
+        const sources = retrieval.chunks.map((chunk) => ({
+            path: chunk.sourcePath,
+            heading: chunk.heading,
+            relevanceScore: chunk.relevanceScore,
+            matchType: chunk.matchType,
+        }));
+        this.emit({
+            type: 'generate:complete',
+            answerLength: generateResult.answer.length,
+            citationCount: sources.length,
+            durationMs: Date.now() - routeStart,
+            timestamp: Date.now(),
+        });
+        // --- Assemble final result ---
+        const totalDuration = Date.now() - routeStart;
+        // Surface capability recommendations from the execution plan into the
+        // QueryResult so consumers (bots, CLI, API) can access them directly.
+        const recommendations = executionPlan.skills.length > 0 ||
+            executionPlan.tools.length > 0 ||
+            executionPlan.extensions.length > 0
+            ? {
+                skills: executionPlan.skills.map((s) => ({
+                    skillId: s.skillId,
+                    reasoning: s.reasoning,
+                    confidence: s.confidence,
+                })),
+                tools: executionPlan.tools.map((t) => ({
+                    toolId: t.toolId,
+                    reasoning: t.reasoning,
+                    confidence: t.confidence,
+                })),
+                extensions: executionPlan.extensions.map((e) => ({
+                    extensionId: e.extensionId,
+                    reasoning: e.reasoning,
+                    confidence: e.confidence,
+                })),
+            }
+            : undefined;
+        const grounding = await this.verifyGeneratedCitations(generateResult.answer, retrieval.chunks);
+        const result = {
+            answer: generateResult.answer,
+            classification,
+            sources,
+            researchSynthesis: retrieval.researchSynthesis,
+            durationMs: totalDuration,
+            tiersUsed,
+            fallbacksUsed,
+            grounding,
+            recommendations,
+        };
+        if (this.config.cacheResults !== false) {
+            this.routeResultCache.set(cacheKey, this.cloneQueryResult(result));
+        }
+        // Log recommendations in verbose mode for observability
+        if (result.recommendations) {
+            const r = result.recommendations;
+            const parts = [];
+            if (r.skills.length)
+                parts.push(`skills: ${r.skills.map((s) => s.skillId).join(', ')}`);
+            if (r.tools.length)
+                parts.push(`tools: ${r.tools.map((t) => t.toolId).join(', ')}`);
+            if (r.extensions.length)
+                parts.push(`extensions: ${r.extensions.map((e) => e.extensionId).join(', ')}`);
+            if (parts.length)
+                console.log(`[QueryRouter] recommendations | ${parts.join(' | ')}`);
+        }
+        this.emit({
+            type: 'route:complete',
+            result,
+            durationMs: totalDuration,
+            timestamp: Date.now(),
+        });
+        return result;
+    }
+    /**
+     * Tear down resources and release references.
+     *
+     * Shuts down embedding and vector store managers if they were initialised,
+     * then nulls out all component references. Safe to call multiple times.
+     * After close(), the router must be re-initialised via {@link init} before
+     * further use.
+     */
+    async close() {
+        // Shut down RAG modules if they were initialised
+        try {
+            if (this.embeddingManager && typeof this.embeddingManager.shutdown === 'function') {
+                await this.embeddingManager.shutdown();
+            }
+        }
+        catch { /* best-effort cleanup */ }
+        try {
+            if (this.vectorStoreManager) {
+                await this.vectorStoreManager.shutdownAllProviders();
+            }
+        }
+        catch { /* best-effort cleanup */ }
+        try {
+            if (this.providerManager) {
+                await this.providerManager.shutdown();
+            }
+        }
+        catch { /* best-effort cleanup */ }
+        this.embeddingManager = null;
+        this.vectorStoreManager = null;
+        this.providerManager = null;
+        this.embeddingDimension = 0;
+        this.embeddingStatus = 'disabled-no-key';
+        this.classifier = null;
+        this.dispatcher = null;
+        this.generator = null;
+        this.keywordFallback = null;
+        this.corpus = [];
+        this.topics = [];
+        this.events = [];
+        this.clearRouteResultCache();
+        this.initialized = false;
+    }
+    /**
+     * Return lightweight corpus/index stats for observability and host startup
+     * logs.
+     *
+     * Useful after {@link init} so callers can confirm the router loaded a real
+     * corpus instead of only knowing that initialisation completed.
+     */
+    getCorpusStats() {
+        const vectorActive = Boolean(this.embeddingManager && this.vectorStoreManager);
+        const platformKnowledge = this.getPlatformKnowledgeCounts();
+        const graphRuntimeMode = this.config.graphEnabled
+            ? (this.hasLiveGraphRuntime() ? 'active' : this.hasHeuristicGraphRuntime() ? 'heuristic' : 'placeholder')
+            : 'disabled';
+        const deepResearchRuntimeMode = this.config.deepResearchEnabled
+            ? (this.hasLiveDeepResearchRuntime()
+                ? 'active'
+                : this.hasHeuristicDeepResearchRuntime()
+                    ? 'heuristic'
+                    : 'placeholder')
+            : 'disabled';
+        return {
+            initialized: this.initialized,
+            configuredPathCount: this.config.knowledgeCorpus.length,
+            chunkCount: this.corpus.length,
+            topicCount: this.topics.length,
+            sourceCount: new Set(this.corpus.map((chunk) => chunk.sourcePath)).size,
+            platformKnowledge,
+            retrievalMode: vectorActive ? 'vector+keyword-fallback' : 'keyword-only',
+            embeddingStatus: vectorActive ? 'active' : this.embeddingStatus,
+            embeddingDimension: vectorActive ? this.embeddingDimension : 0,
+            graphEnabled: this.config.graphEnabled,
+            deepResearchEnabled: this.config.deepResearchEnabled,
+            graphRuntimeMode,
+            rerankRuntimeMode: this.hasLiveRerankerRuntime()
+                ? 'active'
+                : this.hasHeuristicRerankerRuntime()
+                    ? 'heuristic'
+                    : 'placeholder',
+            deepResearchRuntimeMode,
+        };
+    }
+    getPlatformKnowledgeCounts() {
+        const counts = {
+            total: 0,
+            tools: 0,
+            skills: 0,
+            faq: 0,
+            api: 0,
+            troubleshooting: 0,
+        };
+        for (const chunk of this.corpus) {
+            if (!chunk.sourcePath.startsWith('platform:'))
+                continue;
+            counts.total += 1;
+            if (chunk.sourcePath.startsWith('platform:tools/'))
+                counts.tools += 1;
+            else if (chunk.sourcePath.startsWith('platform:skills/'))
+                counts.skills += 1;
+            else if (chunk.sourcePath.startsWith('platform:faq/'))
+                counts.faq += 1;
+            else if (chunk.sourcePath.startsWith('platform:api/'))
+                counts.api += 1;
+            else if (chunk.sourcePath.startsWith('platform:troubleshooting/'))
+                counts.troubleshooting += 1;
+        }
+        return counts;
+    }
+    // ==========================================================================
+    // PRIVATE — Corpus loading
+    // ==========================================================================
+    /**
+     * Load and chunk markdown files from the configured corpus directories.
+     *
+     * Recursively walks each directory, reads .md and .mdx files, and splits
+     * their content by h1-h3 headings. Each heading section becomes a
+     * CorpusChunk (capped at {@link MAX_CHUNK_CHARS} characters, minimum
+     * {@link MIN_CHUNK_CHARS} to filter out trivially small sections).
+     *
+     * @param paths - Array of directory paths to scan for markdown files.
+     * @returns Array of CorpusChunk objects ready for indexing.
+     */
+    loadCorpus(paths) {
+        const chunks = [];
+        let chunkIndex = 0;
+        for (const dirPath of paths) {
+            if (!existsSync(dirPath)) {
+                continue;
+            }
+            this.walkDir(dirPath, (filePath) => {
+                const ext = extname(filePath);
+                if (!MARKDOWN_EXTENSIONS.has(ext)) {
+                    return;
+                }
+                try {
+                    const content = readFileSync(filePath, 'utf-8');
+                    const sections = this.splitByHeadings(content, filePath);
+                    for (const section of sections) {
+                        if (section.content.length < MIN_CHUNK_CHARS) {
+                            continue;
+                        }
+                        chunks.push({
+                            id: `chunk_${chunkIndex++}`,
+                            heading: section.heading,
+                            content: section.content.slice(0, MAX_CHUNK_CHARS),
+                            sourcePath: filePath,
+                        });
+                    }
+                }
+                catch {
+                    // Skip unreadable files gracefully
+                }
+            });
+        }
+        return chunks;
+    }
+    /**
+     * Build a clear init-time error for empty or unreadable corpora.
+     *
+     * The router can technically operate with keyword fallback only, but it
+     * should not silently mark itself ready when no corpus content was loaded
+     * at all. Callers usually interpret a successful `init()` as "docs loaded".
+     *
+     * @param paths - Configured knowledge corpus directory paths.
+     * @returns Human-readable error message for throwing from {@link init}.
+     */
+    buildEmptyCorpusError(paths) {
+        return ('QueryRouter init failed: no readable markdown corpus chunks were loaded. ' +
+            `Checked paths: ${paths.join(', ')}. ` +
+            'Make sure at least one configured directory exists and contains readable ' +
+            '.md or .mdx files with non-trivial section content.');
+    }
+    /**
+     * Load the bundled platform knowledge corpus that ships with @framers/agentos.
+     *
+     * The corpus file (`knowledge/platform-corpus.json`) is generated at build
+     * time by `scripts/build-knowledge-corpus.mjs` and contains tool reference
+     * entries, skill summaries, FAQ, API reference, and troubleshooting guides.
+     *
+     * Falls back gracefully if the file is missing (e.g., in development before
+     * the knowledge build step has run).
+     *
+     * @returns Loaded platform corpus chunks, or empty array if unavailable.
+     */
+    loadPlatformKnowledge() {
+        const candidates = [
+            // Published package layout: knowledge/ sits next to dist/
+            join(MODULE_DIR, '../../knowledge/platform-corpus.json'),
+            // Source layout: knowledge/ sits at package root, src/ is one level down
+            join(MODULE_DIR, '../../../knowledge/platform-corpus.json'),
+        ];
+        for (const corpusPath of candidates) {
+            if (!existsSync(corpusPath))
+                continue;
+            try {
+                const raw = readFileSync(corpusPath, 'utf-8');
+                const entries = JSON.parse(raw);
+                const chunks = entries.map((entry) => ({
+                    id: entry.id,
+                    heading: entry.heading,
+                    content: entry.content,
+                    sourcePath: `platform:${entry.category}/${entry.id}`,
+                }));
+                console.log(`[QueryRouter] Loaded ${chunks.length} platform knowledge entries`);
+                return chunks;
+            }
+            catch {
+                // Platform corpus not parseable — skip silently
+            }
+        }
+        return [];
+    }
+    /**
+     * Recursively walk a directory tree, invoking a callback for each file.
+     *
+     * @param dir - The directory to walk.
+     * @param callback - Function called with the absolute path of each file.
+     */
+    walkDir(dir, callback) {
+        try {
+            const entries = readdirSync(dir, { withFileTypes: true });
+            for (const entry of entries) {
+                const fullPath = join(dir, entry.name);
+                if (entry.isDirectory()) {
+                    this.walkDir(fullPath, callback);
+                }
+                else if (entry.isFile()) {
+                    callback(fullPath);
+                }
+            }
+        }
+        catch {
+            // Skip inaccessible directories gracefully
+        }
+    }
+    /**
+     * Split markdown content into sections by h1-h3 headings.
+     *
+     * Each section captures the heading text and the content between it and
+     * the next heading (or end of file). Content before the first heading is
+     * assigned a heading of "(intro)".
+     *
+     * @param content - The raw markdown file content.
+     * @param sourcePath - File path (used for the section's sourcePath field).
+     * @returns Array of sections with heading and content fields.
+     */
+    splitByHeadings(content, sourcePath) {
+        const lines = content.split('\n');
+        const sections = [];
+        let currentHeading = '(intro)';
+        let currentLines = [];
+        for (const line of lines) {
+            const match = line.match(HEADING_REGEX);
+            if (match) {
+                // Flush the previous section
+                if (currentLines.length > 0) {
+                    sections.push({
+                        heading: currentHeading,
+                        content: currentLines.join('\n').trim(),
+                    });
+                }
+                currentHeading = match[1].trim();
+                currentLines = [];
+            }
+            else {
+                currentLines.push(line);
+            }
+        }
+        // Flush the final section
+        if (currentLines.length > 0) {
+            sections.push({
+                heading: currentHeading,
+                content: currentLines.join('\n').trim(),
+            });
+        }
+        return sections;
+    }
+    // ==========================================================================
+    // PRIVATE — Corpus embedding
+    // ==========================================================================
+    /**
+     * Embed all loaded corpus chunks into the vector store using real
+     * EmbeddingManager and VectorStoreManager instances.
+     *
+     * The method dynamically imports the RAG modules to keep them optional —
+     * if the imports fail or initialisation fails (e.g., no API key), the error
+     * is caught and logged as a warning. The router will continue to function
+     * using the KeywordFallback engine for all retrieval.
+     *
+     * Steps:
+     * 1. Dynamic-import AIModelProviderManager, EmbeddingManager, VectorStoreManager
+     * 2. Initialise the provider manager with the configured embedding provider
+     * 3. Initialise the embedding manager with the configured model
+     * 4. Initialise the vector store manager with an in-memory provider
+     * 5. Create a collection with the correct dimension
+     * 6. Embed all corpus chunks in batches of 50
+     * 7. Upsert the resulting VectorDocuments into the vector store
+     * 8. Cache embeddings on CorpusChunk.embedding for potential reuse
+     */
+    async embedCorpus() {
+        if (this.corpus.length === 0) {
+            return;
+        }
+        // Quick check: bail out early if there's obviously no API key configured.
+        // This avoids the overhead of dynamic imports and provider initialization
+        // in test environments and when no embedding provider is available.
+        const embeddingApiKey = this.getEmbeddingApiKey();
+        if (!embeddingApiKey) {
+            this.disableVectorRetrieval('disabled-no-key');
+            console.debug('[QueryRouter] No embedding API key configured; skipping vector store embedding (keyword fallback active).');
+            return;
+        }
+        try {
+            // --- Dynamic imports to keep RAG modules optional ---
+            const [{ AIModelProviderManager: AIModelProviderManagerClass }, { EmbeddingManager: EmbeddingManagerClass }, { VectorStoreManager: VectorStoreManagerClass },] = await Promise.all([
+                import('../../../core/llm/providers/AIModelProviderManager.js'),
+                import('../../../cognition/rag/EmbeddingManager.js'),
+                import('../../../cognition/rag/VectorStoreManager.js'),
+            ]);
+            // --- 1. Initialise the AI model provider manager ---
+            const pm = new AIModelProviderManagerClass();
+            await pm.initialize({
+                providers: [
+                    {
+                        providerId: this.config.embeddingProvider,
+                        enabled: true,
+                        config: {
+                            apiKey: embeddingApiKey,
+                            ...(this.getEmbeddingBaseUrl() ? { baseUrl: this.getEmbeddingBaseUrl() } : {}),
+                        },
+                        isDefault: true,
+                    },
+                ],
+            });
+            this.providerManager = pm;
+            // --- 2. Initialise the embedding manager ---
+            const em = new EmbeddingManagerClass();
+            const embeddingModelId = this.config.embeddingModel;
+            const embeddingProviderId = this.config.embeddingProvider;
+            // Determine dimension: use a known dimension for common models, or
+            // try to derive it by generating a single test embedding.
+            let dimension = this.getKnownDimension(embeddingModelId);
+            await em.initialize({
+                embeddingModels: [
+                    {
+                        modelId: embeddingModelId,
+                        providerId: embeddingProviderId,
+                        dimension: dimension || 1536, // initial guess; corrected below if needed
+                        isDefault: true,
+                    },
+                ],
+                defaultModelId: embeddingModelId,
+                defaultBatchSize: 50,
+            }, pm);
+            this.embeddingManager = em;
+            // If dimension was unknown, generate a probe embedding to discover it
+            if (!dimension) {
+                const probe = await em.generateEmbeddings({ texts: ['dimension probe'] });
+                if (probe.embeddings.length > 0 && probe.embeddings[0].length > 0) {
+                    dimension = probe.embeddings[0].length;
+                }
+                else {
+                    dimension = 1536; // safe fallback for OpenAI models
+                }
+            }
+            this.embeddingDimension = dimension;
+            this.embeddingStatus = 'active';
+            // --- 3. Initialise the vector store manager (in-memory) ---
+            const vsm = new VectorStoreManagerClass();
+            const collectionName = this.corpusDataSourceId;
+            await vsm.initialize({
+                managerId: 'query-router-vsm',
+                providers: [{ id: 'mem', type: 'in_memory' }],
+                defaultProviderId: 'mem',
+            }, [
+                {
+                    dataSourceId: collectionName,
+                    displayName: 'QueryRouter Corpus',
+                    vectorStoreProviderId: 'mem',
+                    actualNameInProvider: collectionName,
+                    embeddingDimension: dimension,
+                },
+            ]);
+            this.vectorStoreManager = vsm;
+            // --- 4. Create the collection ---
+            const { store, collectionName: resolvedName } = await vsm.getStoreForDataSource(collectionName);
+            if (typeof store.createCollection === 'function') {
+                await store.createCollection(resolvedName, dimension);
+            }
+            // --- 5. Embed corpus chunks in batches of 50 ---
+            const BATCH_SIZE = 50;
+            const allDocuments = [];
+            for (let i = 0; i < this.corpus.length; i += BATCH_SIZE) {
+                const batch = this.corpus.slice(i, i + BATCH_SIZE);
+                const texts = batch.map((c) => c.content);
+                const result = await em.generateEmbeddings({ texts });
+                for (let j = 0; j < batch.length; j++) {
+                    const embedding = result.embeddings[j];
+                    if (!embedding || embedding.length === 0) {
+                        continue; // skip chunks that failed to embed
+                    }
+                    // Cache embedding on the CorpusChunk for potential later reuse
+                    batch[j].embedding = embedding;
+                    allDocuments.push({
+                        id: batch[j].id,
+                        embedding,
+                        textContent: batch[j].content,
+                        metadata: {
+                            heading: batch[j].heading,
+                            sourcePath: batch[j].sourcePath,
+                        },
+                    });
+                }
+            }
+            // --- 6. Upsert into vector store ---
+            if (allDocuments.length > 0) {
+                await store.upsert(resolvedName, allDocuments);
+            }
+            console.log(`[QueryRouter] Embedded ${allDocuments.length} chunks into vector store (dim=${dimension})`);
+        }
+        catch (error) {
+            // Non-fatal: warn and continue — keyword fallback still works
+            const message = error instanceof Error ? error.message : String(error);
+            console.warn(`[QueryRouter] Embedding initialisation failed, falling back to keyword search: ${message}`);
+            this.disableVectorRetrieval('failed-init');
+        }
+    }
+    async syncIndexedCorpusChunks(chunks) {
+        if (chunks.length === 0) {
+            return;
+        }
+        const appendedChunks = this.appendCorpusChunks(chunks);
+        await this.indexAdditionalCorpusChunks(appendedChunks);
+        this.rebuildCorpusSearchState();
+    }
+    appendCorpusChunks(chunks) {
+        const existingKeys = new Set(this.corpus.map((chunk) => `${chunk.sourcePath}\u0000${chunk.heading}\u0000${chunk.content}`));
+        const filteredChunks = chunks
+            .map((chunk) => ({
+            heading: chunk.heading,
+            sourcePath: chunk.sourcePath,
+            content: chunk.content.slice(0, MAX_CHUNK_CHARS).trim(),
+        }))
+            .filter((chunk) => {
+            if (chunk.content.length < MIN_CHUNK_CHARS) {
+                return false;
+            }
+            const dedupeKey = `${chunk.sourcePath}\u0000${chunk.heading}\u0000${chunk.content}`;
+            if (existingKeys.has(dedupeKey)) {
+                return false;
+            }
+            existingKeys.add(dedupeKey);
+            return true;
+        });
+        const startIndex = this.corpus.length;
+        const appendedChunks = filteredChunks.map((chunk, index) => ({
+            id: `gh_${startIndex + index}`,
+            heading: chunk.heading,
+            content: chunk.content,
+            sourcePath: chunk.sourcePath,
+        }));
+        this.corpus.push(...appendedChunks);
+        return appendedChunks;
+    }
+    rebuildCorpusSearchState() {
+        this.keywordFallback = new KeywordFallback(this.corpus);
+        const topicExtractor = new TopicExtractor();
+        this.topics = topicExtractor.extract(this.corpus);
+        if (this.classifier) {
+            this.classifier = this.createClassifier(topicExtractor.formatForPrompt(this.topics));
+        }
+        this.clearRouteResultCache();
+    }
+    createClassifier(topicList) {
+        const classifier = new QueryClassifier({
+            model: this.config.classifierModel,
+            provider: this.config.classifierProvider,
+            confidenceThreshold: this.config.confidenceThreshold,
+            maxTier: this.config.maxTier,
+            topicList,
+            toolList: this.formatToolList(this.config.availableTools),
+            apiKey: this.getLlmApiKey(),
+            baseUrl: this.getLlmBaseUrl(),
+        });
+        if (this.capabilityDiscoveryEngine) {
+            classifier.setCapabilityDiscoveryEngine(this.capabilityDiscoveryEngine);
+        }
+        return classifier;
+    }
+    trimConversationHistory(conversationHistory) {
+        return conversationHistory?.slice(-this.config.conversationWindowSize);
+    }
+    buildRouteCacheKey(query, conversationHistory, options) {
+        return JSON.stringify({
+            query,
+            conversationHistory: conversationHistory ?? null,
+            excludedCapabilityIds: [...(options?.excludedCapabilityIds ?? [])].sort(),
+            hasUnifiedRetriever: Boolean(this.unifiedRetriever),
+            verifyCitations: this.config.verifyCitations === true,
+        });
+    }
+    clearRouteResultCache() {
+        this.routeResultCache.clear();
+    }
+    cloneQueryResult(result) {
+        if (typeof structuredClone === 'function') {
+            return structuredClone(result);
+        }
+        return JSON.parse(JSON.stringify(result));
+    }
+    async verifyGeneratedCitations(answer, chunks) {
+        if (this.config.verifyCitations !== true) {
+            return undefined;
+        }
+        if (!this.embeddingManager || chunks.length === 0) {
+            return undefined;
+        }
+        const sources = chunks
+            .filter((chunk) => chunk.content.trim().length > 0)
+            .map((chunk) => ({
+            content: chunk.content,
+            title: chunk.heading,
+            url: chunk.sourcePath,
+        }));
+        if (sources.length === 0) {
+            return undefined;
+        }
+        try {
+            const verifier = new CitationVerifier({
+                embedFn: async (texts) => {
+                    const result = await this.embeddingManager.generateEmbeddings({ texts });
+                    return result.embeddings;
+                },
+            });
+            return await verifier.verify(answer, sources);
+        }
+        catch {
+            return undefined;
+        }
+    }
+    async indexAdditionalCorpusChunks(chunks) {
+        if (chunks.length === 0 ||
+            !this.embeddingManager ||
+            !this.vectorStoreManager ||
+            this.embeddingStatus !== 'active') {
+            return;
+        }
+        try {
+            const { store, collectionName } = await this.vectorStoreManager.getStoreForDataSource(this.corpusDataSourceId);
+            const BATCH_SIZE = 50;
+            const allDocuments = [];
+            for (let i = 0; i < chunks.length; i += BATCH_SIZE) {
+                const batch = chunks.slice(i, i + BATCH_SIZE);
+                const result = await this.embeddingManager.generateEmbeddings({
+                    texts: batch.map((chunk) => chunk.content),
+                });
+                for (let j = 0; j < batch.length; j++) {
+                    const embedding = result.embeddings[j];
+                    if (!embedding || embedding.length === 0) {
+                        continue;
+                    }
+                    batch[j].embedding = embedding;
+                    allDocuments.push({
+                        id: batch[j].id,
+                        embedding,
+                        textContent: batch[j].content,
+                        metadata: {
+                            heading: batch[j].heading,
+                            sourcePath: batch[j].sourcePath,
+                        },
+                    });
+                }
+            }
+            if (allDocuments.length > 0) {
+                await store.upsert(collectionName, allDocuments);
+            }
+        }
+        catch (error) {
+            const message = error instanceof Error ? error.message : String(error);
+            console.warn(`[QueryRouter] Incremental GitHub corpus embedding failed; falling back to keyword search: ${message}`);
+            this.disableVectorRetrieval('failed-init');
+        }
+    }
+    disableVectorRetrieval(status) {
+        this.embeddingManager = null;
+        this.vectorStoreManager = null;
+        this.providerManager = null;
+        this.embeddingDimension = 0;
+        this.embeddingStatus = status;
+    }
+    /**
+     * Resolve API key for embedding calls.
+     * Falls back through embedding-specific → global → env var scan.
+     */
+    getEmbeddingApiKey() {
+        if (this.config.embeddingApiKey)
+            return this.config.embeddingApiKey;
+        // Fall through to general LLM key resolver
+        return this.getLlmApiKey();
+    }
+    getEmbeddingBaseUrl() {
+        if (this.config.embeddingBaseUrl !== undefined) {
+            return this.config.embeddingBaseUrl;
+        }
+        if (this.config.embeddingApiKey !== undefined) {
+            return undefined;
+        }
+        if (this.config.baseUrl !== undefined) {
+            return this.config.baseUrl;
+        }
+        if (this.config.apiKey !== undefined) {
+            return undefined;
+        }
+        if (process.env.OPENAI_API_KEY) {
+            return undefined;
+        }
+        if (process.env.OPENROUTER_API_KEY) {
+            return 'https://openrouter.ai/api/v1';
+        }
+        return undefined;
+    }
+    /**
+     * Resolve API key for LLM calls.
+     *
+     * Checks config override first, then scans all provider env vars in priority
+     * order. Returns empty string for keyless providers (claude-code-cli, gemini-cli)
+     * which is fine — generateText() handles them via CLISubprocessBridge.
+     */
+    getLlmApiKey() {
+        if (this.config.apiKey)
+            return this.config.apiKey;
+        // Check all providers in priority order.
+        // Direct provider keys (OpenAI, Anthropic, etc.) take precedence over
+        // aggregator keys (OpenRouter) so that the native endpoint is used when
+        // both are configured.
+        const envKeys = [
+            'OPENAI_API_KEY',
+            'ANTHROPIC_API_KEY',
+            'GEMINI_API_KEY',
+            'GROQ_API_KEY',
+            'TOGETHER_API_KEY',
+            'MISTRAL_API_KEY',
+            'XAI_API_KEY',
+            'OPENROUTER_API_KEY',
+        ];
+        for (const key of envKeys) {
+            if (process.env[key])
+                return process.env[key];
+        }
+        // Keyless providers (CLI) don't need an API key — empty string is fine
+        return '';
+    }
+    /**
+     * Resolve base URL for LLM calls.
+     *
+     * Only OpenRouter and Ollama need custom base URLs. All other providers
+     * (including CLI) use their default endpoints via generateText() resolution.
+     */
+    getLlmBaseUrl() {
+        if (this.config.baseUrl !== undefined) {
+            return this.config.baseUrl;
+        }
+        // If user provided an explicit API key, don't override the URL
+        if (this.config.apiKey !== undefined) {
+            return undefined;
+        }
+        // Direct provider keys don't need a custom base URL — only fall through
+        // to OpenRouter when no direct key is present.
+        const directKeys = [
+            'OPENAI_API_KEY',
+            'ANTHROPIC_API_KEY',
+            'GEMINI_API_KEY',
+            'GROQ_API_KEY',
+            'TOGETHER_API_KEY',
+            'MISTRAL_API_KEY',
+            'XAI_API_KEY',
+        ];
+        if (directKeys.some((k) => process.env[k])) {
+            return undefined;
+        }
+        // OpenRouter needs a custom base URL
+        if (process.env.OPENROUTER_API_KEY) {
+            return 'https://openrouter.ai/api/v1';
+        }
+        // Ollama needs its base URL
+        if (process.env.OLLAMA_BASE_URL) {
+            return process.env.OLLAMA_BASE_URL;
+        }
+        return undefined;
+    }
+    /**
+     * Return a known embedding dimension for common models.
+     *
+     * This avoids an extra API call when the dimension can be statically
+     * determined from the model identifier.
+     *
+     * @param modelId - The embedding model identifier.
+     * @returns The known dimension, or 0 if unknown.
+     */
+    getKnownDimension(modelId) {
+        const KNOWN_DIMENSIONS = {
+            'text-embedding-3-small': 1536,
+            'text-embedding-3-large': 3072,
+            'text-embedding-ada-002': 1536,
+        };
+        return KNOWN_DIMENSIONS[modelId] ?? 0;
+    }
+    /**
+     * Whether graph expansion is backed by a live implementation.
+     *
+     * Hosts should not treat `graphEnabled` as meaning GraphRAG is actually live.
+     * `active` is reserved for a host-injected or future provider-backed graph
+     * runtime rather than the built-in heuristic.
+     */
+    hasLiveGraphRuntime() {
+        return typeof this.config.graphExpand === 'function';
+    }
+    /**
+     * Whether graph expansion is backed by the built-in heuristic expansion.
+     */
+    hasHeuristicGraphRuntime() {
+        return true;
+    }
+    /**
+     * Whether reranking is backed by a live implementation.
+     *
+     * `active` is reserved for a host-injected or future provider-backed
+     * reranker rather than the built-in lexical heuristic.
+     */
+    hasLiveRerankerRuntime() {
+        return typeof this.config.rerank === 'function';
+    }
+    /**
+     * Whether reranking is backed by the built-in lexical reranker.
+     */
+    hasHeuristicRerankerRuntime() {
+        return true;
+    }
+    /**
+     * Whether deep research is backed by a live implementation.
+     *
+     * `deepResearchEnabled` only means the branch may be attempted by config.
+     * `active` is reserved for a host-injected or future provider-backed
+     * research runtime rather than the built-in local-corpus heuristic.
+     */
+    hasLiveDeepResearchRuntime() {
+        return typeof this.config.deepResearch === 'function';
+    }
+    /**
+     * Whether deep research is backed by the built-in corpus-only heuristic.
+     */
+    hasHeuristicDeepResearchRuntime() {
+        return true;
+    }
+    // ==========================================================================
+    // PRIVATE — Retrieval callbacks (injected into QueryDispatcher)
+    // ==========================================================================
+    /**
+     * HyDE (Hypothetical Document Embeddings) search callback for the dispatcher.
+     *
+     * Generates a hypothetical answer to the query using the LLM, then searches
+     * for documents similar to that hypothetical answer. Falls back to standard
+     * vector search if no generation provider is available.
+     *
+     * @param query - The user's query string.
+     * @param topK - Maximum number of chunks to return.
+     * @returns Promise resolving to an array of matched chunks.
+     */
+    async hydeSearch(query, topK) {
+        // HyDE falls back to standard vector search in the built-in implementation.
+        // A host-injected HyDE retriever would generate a hypothetical document first.
+        return this.vectorSearch(query, topK);
+    }
+    /**
+     * Query decomposition callback for the dispatcher.
+     *
+     * Splits a complex multi-part query into independent sub-queries.
+     * The built-in implementation uses simple sentence splitting as a heuristic.
+     * A host-injected decomposer would use an LLM for semantic decomposition.
+     *
+     * @param query - The original multi-part user query.
+     * @param maxSubQueries - Maximum number of sub-queries to generate.
+     * @returns Array of decomposed sub-query strings.
+     */
+    async decomposeQuery(query, maxSubQueries) {
+        // Built-in heuristic: split on sentence boundaries or question marks.
+        const parts = query
+            .split(/[?.!]\s+/)
+            .map(s => s.trim())
+            .filter(s => s.length > 0);
+        if (parts.length <= 1)
+            return [query];
+        return parts.slice(0, maxSubQueries);
+    }
+    /**
+     * Vector search callback for the dispatcher.
+     *
+     * When the EmbeddingManager and VectorStoreManager are available, this method
+     * embeds the query, queries the vector store, and maps the results to
+     * RetrievedChunk objects. If the RAG modules are not available (e.g., embedding
+     * init failed), it falls back to the KeywordFallback engine and emits a
+     * retrieve:fallback event.
+     *
+     * @param query - The user's query string.
+     * @param topK - Maximum number of chunks to return.
+     * @returns Promise resolving to an array of matched chunks.
+     */
+    async vectorSearch(query, topK) {
+        // --- Real vector search when RAG modules are available ---
+        if (this.embeddingManager && this.vectorStoreManager) {
+            try {
+                // Embed the query
+                const queryResult = await this.embeddingManager.generateEmbeddings({
+                    texts: [query],
+                });
+                const queryEmbedding = queryResult.embeddings[0];
+                if (!queryEmbedding || queryEmbedding.length === 0) {
+                    // Embedding failed for the query — fall through to keyword fallback
+                    throw new Error('Query embedding returned empty vector');
+                }
+                // Query the vector store
+                const { store, collectionName } = await this.vectorStoreManager.getStoreForDataSource(this.corpusDataSourceId);
+                const searchResults = await store.query(collectionName, queryEmbedding, {
+                    topK,
+                    includeTextContent: true,
+                    includeMetadata: true,
+                });
+                // Map retrieved vector documents to RetrievedChunk[]
+                return searchResults.documents.map((doc) => ({
+                    id: doc.id,
+                    content: doc.textContent ?? '',
+                    heading: doc.metadata?.heading ?? '',
+                    sourcePath: doc.metadata?.sourcePath ?? '',
+                    relevanceScore: doc.similarityScore ?? 0,
+                    matchType: 'vector',
+                }));
+            }
+            catch (error) {
+                // On any error during vector search, fall back to keyword search
+                const message = error instanceof Error ? error.message : String(error);
+                console.warn(`[QueryRouter] Vector search failed, falling back to keyword search: ${message}`);
+                this.emit({
+                    type: 'retrieve:fallback',
+                    strategy: 'keyword-fallback',
+                    reason: `Vector search error: ${message}`,
+                    timestamp: Date.now(),
+                });
+            }
+        }
+        else {
+            // RAG modules not available — emit a fallback event
+            this.emit({
+                type: 'retrieve:fallback',
+                strategy: 'keyword-fallback',
+                reason: 'Embeddings unavailable; using keyword search',
+                timestamp: Date.now(),
+            });
+        }
+        // --- Keyword fallback ---
+        if (!this.keywordFallback) {
+            return [];
+        }
+        return this.keywordFallback.search(query, topK);
+    }
+    /**
+     * Graph expansion callback for the dispatcher.
+     *
+     * Built-in heuristic graph expansion over the loaded corpus.
+     *
+     * This is not yet a true GraphRAG engine. It expands from seed chunks by
+     * preferring:
+     * - chunks from the same source document
+     * - heading overlap with seed headings/content
+     * - content overlap with seed headings/content
+     *
+     * @param seeds - Seed chunks to expand from.
+     * @returns Promise resolving to related chunks marked as `graph`.
+     */
+    async graphExpand(seeds) {
+        if (seeds.length === 0 || this.corpus.length === 0) {
+            return [];
+        }
+        const seedIds = new Set(seeds.map((seed) => seed.id));
+        const seedSourcePaths = new Set(seeds.map((seed) => seed.sourcePath));
+        const seedTerms = new Set();
+        for (const seed of seeds) {
+            for (const term of this.tokenizeForRerank(`${seed.heading} ${seed.content}`)) {
+                seedTerms.add(term);
+            }
+        }
+        const scored = this.corpus
+            .filter((chunk) => !seedIds.has(chunk.id))
+            .map((chunk, index) => {
+            const sameSourceBoost = seedSourcePaths.has(chunk.sourcePath) ? 0.48 : 0;
+            const headingOverlap = this.computeTermOverlap(seedTerms, this.tokenizeForRerank(chunk.heading));
+            const contentOverlap = this.computeTermOverlap(seedTerms, this.tokenizeForRerank(chunk.content));
+            const score = sameSourceBoost + headingOverlap * 0.32 + contentOverlap * 0.2;
+            return { chunk, index, score };
+        })
+            .filter((entry) => entry.score >= 0.18);
+        scored.sort((left, right) => {
+            if (right.score !== left.score) {
+                return right.score - left.score;
+            }
+            return left.index - right.index;
+        });
+        return scored.slice(0, 8).map(({ chunk, score }) => ({
+            id: chunk.id,
+            heading: chunk.heading,
+            content: chunk.content,
+            sourcePath: chunk.sourcePath,
+            relevanceScore: Math.min(0.99, score),
+            matchType: 'graph',
+        }));
+    }
+    /**
+     * Reranking callback for the dispatcher.
+     *
+     * Built-in heuristic reranker.
+     *
+     * This is not yet a cross-encoder. It reorders candidate chunks by combining:
+     * - original retrieval score
+     * - heading term overlap
+     * - content term overlap
+     * - exact phrase containment
+     *
+     * This gives tier-2 routing a real second-pass ranking step today without
+     * pretending the deeper reranker service is already wired.
+     *
+     * @param query - The user's query.
+     * @param chunks - Candidate chunks to rerank.
+     * @param topN - Maximum number of chunks to keep.
+     * @returns Promise resolving to the best-ranked chunks.
+     */
+    async rerank(query, chunks, topN) {
+        if (chunks.length <= 1) {
+            return chunks.slice(0, topN);
+        }
+        const queryTerms = this.tokenizeForRerank(query);
+        const normalizedQuery = this.normalizeForRerank(query);
+        const scored = chunks.map((chunk, index) => {
+            const headingTerms = this.tokenizeForRerank(chunk.heading);
+            const contentTerms = this.tokenizeForRerank(chunk.content);
+            const headingOverlap = this.computeTermOverlap(queryTerms, headingTerms);
+            const contentOverlap = this.computeTermOverlap(queryTerms, contentTerms);
+            const normalizedText = this.normalizeForRerank(`${chunk.heading} ${chunk.content}`);
+            const exactPhraseBoost = normalizedQuery.length >= 4 && normalizedText.includes(normalizedQuery) ? 0.2 : 0;
+            const score = chunk.relevanceScore * 0.55 +
+                headingOverlap * 0.25 +
+                contentOverlap * 0.2 +
+                exactPhraseBoost;
+            return { chunk, index, score };
+        });
+        scored.sort((left, right) => {
+            if (right.score !== left.score) {
+                return right.score - left.score;
+            }
+            if (right.chunk.relevanceScore !== left.chunk.relevanceScore) {
+                return right.chunk.relevanceScore - left.chunk.relevanceScore;
+            }
+            return left.index - right.index;
+        });
+        return scored.slice(0, topN).map((entry) => entry.chunk);
+    }
+    /**
+     * Deep research callback for the dispatcher.
+     *
+     * Built-in corpus-only research heuristic.
+     *
+     * This is not web-backed research. It runs a few local keyword-based passes
+     * over the loaded corpus using slightly different query formulations, merges
+     * the results, and returns a compact synthesis built from the top findings.
+     *
+     * @param query - The user's query.
+     * @param sources - Normalized research-source hints used to broaden local
+     *                  matching.
+     * @returns Promise resolving to synthesized local-corpus findings.
+     */
+    async deepResearch(query, sources) {
+        if (!this.keywordFallback || this.corpus.length === 0) {
+            return { synthesis: '', sources: [] };
+        }
+        const normalizedSourceHints = sources
+            .map((source) => this.normalizeForRerank(source))
+            .filter(Boolean);
+        const queryTerms = [...this.tokenizeForRerank(query)];
+        const narrowedTerms = queryTerms.slice(0, 4).join(' ');
+        const researchQueries = [
+            query,
+            [query, normalizedSourceHints.join(' ')].filter(Boolean).join(' ').trim(),
+            [query, 'architecture tradeoffs details'].join(' ').trim(),
+            narrowedTerms,
+        ].filter(Boolean);
+        const merged = new Map();
+        for (const researchQuery of researchQueries) {
+            const hits = this.keywordFallback.search(researchQuery, 4);
+            for (const hit of hits) {
+                const existing = merged.get(hit.id);
+                const researchChunk = {
+                    ...hit,
+                    relevanceScore: Math.min(0.99, hit.relevanceScore),
+                    matchType: 'research',
+                };
+                if (!existing || researchChunk.relevanceScore > existing.relevanceScore) {
+                    merged.set(researchChunk.id, researchChunk);
+                }
+            }
+        }
+        const researchChunks = [...merged.values()]
+            .sort((left, right) => right.relevanceScore - left.relevanceScore)
+            .slice(0, 5);
+        if (researchChunks.length === 0) {
+            return { synthesis: '', sources: [] };
+        }
+        const synthesis = researchChunks
+            .map((chunk, index) => `${index + 1}. ${chunk.heading}: ${this.firstSentence(chunk.content)}`)
+            .join('\n');
+        return { synthesis, sources: researchChunks };
+    }
+    /**
+     * Format available tools for the classifier prompt.
+     */
+    formatToolList(availableTools) {
+        return availableTools.length > 0 ? availableTools.join(', ') : '(none available)';
+    }
+    /**
+     * Normalize text for simple lexical reranking.
+     */
+    normalizeForRerank(text) {
+        return text.toLowerCase().replace(/[^a-z0-9]+/g, ' ').trim();
+    }
+    /**
+     * Tokenize text for lexical reranking.
+     */
+    tokenizeForRerank(text) {
+        const STOP_WORDS = new Set([
+            'a',
+            'an',
+            'and',
+            'are',
+            'as',
+            'at',
+            'be',
+            'by',
+            'for',
+            'from',
+            'how',
+            'in',
+            'is',
+            'it',
+            'of',
+            'on',
+            'or',
+            'that',
+            'the',
+            'this',
+            'to',
+            'what',
+            'when',
+            'where',
+            'which',
+            'why',
+            'with',
+        ]);
+        return new Set(this.normalizeForRerank(text)
+            .split(/\s+/)
+            .filter((term) => term.length >= 2 && !STOP_WORDS.has(term)));
+    }
+    /**
+     * Compute overlap ratio between query terms and candidate terms.
+     */
+    computeTermOverlap(queryTerms, candidateTerms) {
+        if (queryTerms.size === 0 || candidateTerms.size === 0) {
+            return 0;
+        }
+        let matches = 0;
+        for (const term of queryTerms) {
+            if (candidateTerms.has(term)) {
+                matches += 1;
+            }
+        }
+        return matches / queryTerms.size;
+    }
+    /**
+     * Extract a short first-sentence style summary from a chunk for synthesis.
+     */
+    firstSentence(text) {
+        const normalized = text.replace(/\s+/g, ' ').trim();
+        const sentence = normalized.match(/^(.{1,220}?[.!?])(?:\s|$)/);
+        return sentence?.[1] ?? normalized.slice(0, 220);
+    }
+    /**
+     * Derive fallback strategy names from classification + retrieval events.
+     */
+    collectFallbacks(classification, events) {
+        const fallbacks = new Set();
+        if (classification.confidence < this.config.confidenceThreshold) {
+            fallbacks.add('low-confidence-classification');
+        }
+        for (const event of events) {
+            if (event.type === 'retrieve:fallback') {
+                fallbacks.add(event.strategy);
+            }
+        }
+        return Array.from(fallbacks);
+    }
+    /**
+     * Approximate the tiers actually exercised when fallback strategies fired.
+     */
+    collectTiersUsed(classification, fallbacksUsed) {
+        const tiers = new Set([classification.tier]);
+        for (const fallback of fallbacksUsed) {
+            if (fallback === 'research-skip') {
+                tiers.add(2);
+            }
+            if (fallback === 'graph-skip' || fallback === 'keyword-fallback' || fallback === 'rerank-skip') {
+                tiers.add(1);
+            }
+        }
+        return Array.from(tiers).sort((a, b) => a - b);
+    }
+    // ==========================================================================
+    // PRIVATE — Event emission
+    // ==========================================================================
+    /**
+     * Store a lifecycle event and emit a structured console log.
+     *
+     * Events are accumulated in the `events` array for later inspection
+     * and also logged to the console with a `[QueryRouter]` prefix for
+     * real-time observability.
+     *
+     * @param event - The typed lifecycle event to emit.
+     */
+    emit(event) {
+        this.events.push(event);
+    }
+    // ==========================================================================
+    // PRIVATE — Guards
+    // ==========================================================================
+    /**
+     * Assert that the router has been initialised via {@link init}.
+     *
+     * @throws Error if `init()` has not been called.
+     */
+    ensureInitialized() {
+        if (!this.initialized) {
+            throw new Error('QueryRouter has not been initialised. Call init() before classify/retrieve/route.');
+        }
+    }
+}
+//# sourceMappingURL=QueryRouter.js.map