mdcontext 0.1.0 → 0.2.0

package/src/embeddings/types.ts

@@ -6,10 +6,43 @@
 // Embedding Provider
 // ============================================================================
 
+export interface BatchProgress {
+  readonly batchIndex: number
+  readonly totalBatches: number
+  readonly processedTexts: number
+  readonly totalTexts: number
+}
+
+export interface EmbedOptions {
+  readonly onBatchProgress?: ((progress: BatchProgress) => void) | undefined
+}
+
 export interface EmbeddingProvider {
   readonly name: string
   readonly dimensions: number
-  embed(texts: string[]): Promise<EmbeddingResult>
+  embed(texts: string[], options?: EmbedOptions): Promise<EmbeddingResult>
+}
+
+/**
+ * Extended embedding provider with metadata about the underlying service.
+ * Implementations like OpenAIProvider include these additional properties.
+ */
+export interface EmbeddingProviderWithMetadata extends EmbeddingProvider {
+  readonly model: string
+  readonly baseURL: string | undefined
+}
+
+/**
+ * Type guard to check if an EmbeddingProvider has extended metadata.
+ * Use this instead of unsafe type casting when accessing model/baseURL.
+ */
+export const hasProviderMetadata = (
+  provider: EmbeddingProvider,
+): provider is EmbeddingProviderWithMetadata => {
+  return (
+    'model' in provider &&
+    typeof (provider as EmbeddingProviderWithMetadata).model === 'string'
+  )
 }
 
 export interface EmbeddingResult {
@@ -33,14 +66,56 @@ export interface VectorEntry {
 export interface VectorIndex {
   readonly version: number
   readonly provider: string
+  readonly providerModel?: string | undefined
+  readonly providerBaseURL?: string | undefined
   readonly dimensions: number
   readonly entries: Record<string, VectorEntry>
   readonly totalCost: number
   readonly totalTokens: number
   readonly createdAt: string
   readonly updatedAt: string
+  /**
+   * HNSW index build parameters (stored for validation on load).
+   * These affect index quality and build time - changes require rebuild.
+   */
+  readonly hnswParams?: HnswIndexParams | undefined
+}
+
+/**
+ * HNSW index parameters stored in metadata.
+ * Used to detect config/index mismatches and recommend rebuilds.
+ */
+export interface HnswIndexParams {
+  /** Max connections per node (M parameter). Default: 16 */
+  readonly m: number
+  /** Construction-time search width. Default: 200 */
+  readonly efConstruction: number
 }
 
+// ============================================================================
+// Quality Modes
+// ============================================================================
+
+/**
+ * Search quality modes for HNSW efSearch parameter.
+ * Higher efSearch values give better recall at the cost of speed.
+ *
+ * - 'fast': efSearch=64, ~40% faster, slight recall reduction
+ * - 'balanced': efSearch=100 (default), good balance
+ * - 'thorough': efSearch=256, ~30% slower, best recall
+ */
+export type SearchQuality = 'fast' | 'balanced' | 'thorough'
+
+/**
+ * efSearch values for each quality mode.
+ * These control the size of the dynamic candidate list during search.
+ */
+export const QUALITY_EF_SEARCH: Record<SearchQuality, number> = {
+  fast: 64,
+  balanced: 100,
+  thorough: 256,
+} as const
+
 // ============================================================================
 // Semantic Search
 // ============================================================================
@@ -52,6 +127,183 @@ export interface SemanticSearchOptions {
   readonly threshold?: number | undefined
   /** Filter by document path pattern */
   readonly pathPattern?: string | undefined
+  /** Search quality mode: fast, balanced (default), or thorough */
+  readonly quality?: SearchQuality | undefined
+  /** Provider configuration override */
+  readonly providerConfig?:
+    | {
+        readonly provider:
+          | 'openai'
+          | 'ollama'
+          | 'lm-studio'
+          | 'openrouter'
+          | 'voyage'
+        readonly baseURL?: string | undefined
+        readonly model?: string | undefined
+      }
+    | undefined
+  /**
+   * Skip query preprocessing (normalize, lowercase, strip punctuation).
+   * Default: false (preprocessing enabled for better recall).
+   * Set to true for exact query matching.
+   */
+  readonly skipPreprocessing?: boolean | undefined
+  /**
+   * Boost results where query terms appear in section headings.
+   * Improves navigation queries like "installation guide" or "API reference".
+   * Default: true (heading boost enabled).
+   */
+  readonly headingBoost?: boolean | undefined
+  /**
+   * Use HyDE (Hypothetical Document Embeddings) for query expansion.
+   * Generates a hypothetical document answering the query using an LLM,
+   * then searches using that document's embedding.
+   *
+   * Best for: complex questions, "how to" queries, ambiguous searches
+   * Adds: ~1-2s latency, LLM API cost
+   * Improvement: 10-30% better recall on complex queries
+   *
+   * Default: false (disabled)
+   */
+  readonly hyde?: boolean | undefined
+  /**
+   * HyDE configuration options (only used when hyde: true).
+   */
+  readonly hydeOptions?:
+    | {
+        /** Model for hypothetical document generation. Default: gpt-4o-mini */
+        readonly model?: string | undefined
+        /** Max tokens for generation. Default: 256 */
+        readonly maxTokens?: number | undefined
+        /** Generation temperature (0-1). Default: 0.3 */
+        readonly temperature?: number | undefined
+      }
+    | undefined
+  /** Lines of context before matches */
+  readonly contextBefore?: number | undefined
+  /** Lines of context after matches */
+  readonly contextAfter?: number | undefined
+}
+
+// ============================================================================
+// Query Preprocessing
+// ============================================================================
+
+// ============================================================================
+// Heading Boost
+// ============================================================================
+
+/** Boost factor per matched term in heading (0.05 = 5% boost per term) */
+const HEADING_BOOST_FACTOR = 0.05
+
+/** Boost factor for important files like README (0.03 = 3% boost) */
+const FILE_IMPORTANCE_BOOST = 0.03
+
+/**
+ * Important file patterns that get ranking boost.
+ * These are typically entry points or high-value documentation.
+ */
+const IMPORTANT_FILE_PATTERNS = [
+  /^readme\.md$/i, // Root README
+  /\/readme\.md$/i, // Nested README
+  /^index\.md$/i, // Index files
+  /\/index\.md$/i,
+  /^getting-?started/i, // Getting started guides
+  /\/getting-?started/i,
+  /^introduction/i, // Introductions
+  /\/introduction/i,
+  /^overview/i, // Overviews
+  /\/overview/i,
+  /^quickstart/i, // Quickstart guides
+  /\/quickstart/i,
+  /^changelog\.md$/i, // Changelogs (useful for "what changed" queries)
+  /\/changelog\.md$/i,
+]
+
+/**
+ * Calculate file importance boost for a search result.
+ * Boosts results from important files like README, index, getting-started.
+ *
+ * @param documentPath - Path to the document
+ * @returns Boost value to add to similarity score (0.0 to 0.03)
+ */
+export const calculateFileImportanceBoost = (documentPath: string): number => {
+  const isImportant = IMPORTANT_FILE_PATTERNS.some((pattern) =>
+    pattern.test(documentPath),
+  )
+  return isImportant ? FILE_IMPORTANCE_BOOST : 0
+}
+
+/**
+ * Calculate heading match boost for a search result.
+ * Boosts results where query terms appear in section headings.
+ *
+ * @param heading - Section heading to check
+ * @param query - Original search query (will be normalized)
+ * @returns Boost value to add to similarity score (0.0 to ~0.15 typically)
+ */
+export const calculateHeadingBoost = (
+  heading: string,
+  query: string,
+): number => {
+  const queryTerms = query.toLowerCase().split(/\s+/).filter(Boolean)
+  if (queryTerms.length === 0) return 0
+
+  const headingLower = heading.toLowerCase()
+  const matchCount = queryTerms.filter((term) =>
+    headingLower.includes(term),
+  ).length
+
+  return matchCount * HEADING_BOOST_FACTOR
+}
+
+/**
+ * Calculate combined ranking boost for a search result.
+ * Combines heading match boost and file importance boost.
+ *
+ * @param heading - Section heading
+ * @param query - Search query
+ * @param documentPath - Path to the document
+ * @returns Combined boost value (0.0 to ~0.18 typically)
+ */
+export const calculateRankingBoost = (
+  heading: string,
+  query: string,
+  documentPath: string,
+): number => {
+  const headingBoost = calculateHeadingBoost(heading, query)
+  const fileBoost = calculateFileImportanceBoost(documentPath)
+  return headingBoost + fileBoost
+}
+
+// ============================================================================
+// Query Preprocessing
+// ============================================================================
+
+/**
+ * Preprocess a search query before embedding to reduce noise and improve recall.
+ *
+ * Transformations applied:
+ * - Convert to lowercase (embeddings are case-insensitive)
+ * - Replace punctuation with spaces (preserves word boundaries)
+ * - Collapse multiple spaces to single space
+ * - Trim leading/trailing whitespace
+ *
+ * This provides 2-5% precision improvement for most queries.
+ *
+ * @param query - Raw search query
+ * @returns Normalized query string
+ */
+export const preprocessQuery = (query: string): string => {
+  return (
+    query
+      .toLowerCase()
+      // Replace punctuation with spaces (preserves word boundaries)
+      .replace(/[^\w\s]/g, ' ')
+      // Collapse multiple spaces
+      .replace(/\s+/g, ' ')
+      .trim()
+  )
 }
 
 export interface SemanticSearchResult {
@@ -60,23 +312,48 @@ export interface SemanticSearchResult {
   readonly heading: string
   readonly similarity: number
   readonly content?: string | undefined
+  /** Context lines with their line numbers (when context is requested) */
+  readonly contextLines?: readonly ContextLine[] | undefined
 }
 
-// ============================================================================
-// Errors
-// ============================================================================
+export interface ContextLine {
+  /** The line number (1-based) */
+  readonly lineNumber: number
+  /** The line text */
+  readonly line: string
+  /**
+   * Whether this line is part of the matched result.
+   *
+   * - For keyword search: true when the line directly matches the query.
+   * - For semantic/hybrid search: true when the line lies within the
+   *   selected/matched section span, even if it is not a direct text match.
+   */
+  readonly isMatch: boolean
+}
 
-export interface EmbedError {
-  readonly _tag: 'EmbedError'
-  readonly cause: 'RateLimit' | 'ApiKey' | 'Network' | 'Unknown'
-  readonly message: string
+/**
+ * Extended semantic search result including metadata about below-threshold results.
+ * Used to provide user feedback when 0 results pass the threshold.
+ */
+export interface SemanticSearchResultWithStats {
+  readonly results: readonly SemanticSearchResult[]
+  /** Number of results found below threshold (only set when includeBelowThresholdStats is true) */
+  readonly belowThresholdCount?: number | undefined
+  /** Highest similarity among below-threshold results */
+  readonly belowThresholdHighest?: number | undefined
+  /** Total results available above threshold before limit was applied */
+  readonly totalAvailable?: number | undefined
 }
 
-export const embedError = (
-  cause: EmbedError['cause'],
-  message: string,
-): EmbedError => ({
-  _tag: 'EmbedError',
-  cause,
-  message,
-})
+// ============================================================================
+// Errors
+// ============================================================================
+// NOTE: Embedding-related errors are defined in src/errors/index.ts:
+// - EmbeddingError: For embedding operation failures (rate limits, quota, network)
+// - ApiKeyMissingError: For missing API keys
+// - ApiKeyInvalidError: For invalid/rejected API keys
+//
+// Use these centralized error types instead of defining errors here.
+// Example:
+//   import { EmbeddingError } from '../errors/index.js'
+//   new EmbeddingError({ reason: 'RateLimit', message: 'Rate limited' })