audrey 0.5.1 → 0.9.0

package/src/recall.js

@@ -1,10 +1,13 @@
-import { computeConfidence, DEFAULT_HALF_LIVES } from './confidence.js';
+import { computeConfidence, DEFAULT_HALF_LIVES, salienceModifier } from './confidence.js';
+import { interferenceModifier } from './interference.js';
+import { contextMatchRatio, contextModifier } from './context.js';
+import { moodCongruenceModifier, affectSimilarity } from './affect.js';
 import { daysBetween, safeJsonParse } from './utils.js';
 
 function computeEpisodicConfidence(ep, now, confidenceConfig = {}) {
   const ageDays = daysBetween(ep.created_at, now);
   const halfLives = confidenceConfig.halfLives || DEFAULT_HALF_LIVES;
-  return computeConfidence({
+  let confidence = computeConfidence({
     sourceType: ep.source,
     supportingCount: 1,
     contradictingCount: 0,
@@ -15,6 +18,8 @@ function computeEpisodicConfidence(ep, now, confidenceConfig = {}) {
     weights: confidenceConfig.weights,
     customSourceReliability: confidenceConfig.sourceReliability,
   });
+  confidence *= salienceModifier(ep.salience);
+  return Math.max(0, Math.min(1, confidence));
 }
 
 function computeSemanticConfidence(sem, now, confidenceConfig = {}) {
@@ -23,7 +28,7 @@ function computeSemanticConfidence(sem, now, confidenceConfig = {}) {
     ? daysBetween(sem.last_reinforced_at, now)
     : ageDays;
   const halfLives = confidenceConfig.halfLives || DEFAULT_HALF_LIVES;
-  return computeConfidence({
+  let confidence = computeConfidence({
     sourceType: 'tool-result',
     supportingCount: sem.supporting_count || 0,
     contradictingCount: sem.contradicting_count || 0,
@@ -34,6 +39,9 @@ function computeSemanticConfidence(sem, now, confidenceConfig = {}) {
     weights: confidenceConfig.weights,
     customSourceReliability: confidenceConfig.sourceReliability,
   });
+  confidence *= interferenceModifier(sem.interference_count || 0, confidenceConfig.interferenceWeight);
+  confidence *= salienceModifier(sem.salience);
+  return Math.max(0, Math.min(1, confidence));
 }
 
 function computeProceduralConfidence(proc, now, confidenceConfig = {}) {
@@ -42,7 +50,7 @@ function computeProceduralConfidence(proc, now, confidenceConfig = {}) {
     ? daysBetween(proc.last_reinforced_at, now)
     : ageDays;
   const halfLives = confidenceConfig.halfLives || DEFAULT_HALF_LIVES;
-  return computeConfidence({
+  let confidence = computeConfidence({
     sourceType: 'tool-result',
     supportingCount: proc.success_count || 0,
     contradictingCount: proc.failure_count || 0,
@@ -53,9 +61,12 @@ function computeProceduralConfidence(proc, now, confidenceConfig = {}) {
     weights: confidenceConfig.weights,
     customSourceReliability: confidenceConfig.sourceReliability,
   });
+  confidence *= interferenceModifier(proc.interference_count || 0, confidenceConfig.interferenceWeight);
+  confidence *= salienceModifier(proc.salience);
+  return Math.max(0, Math.min(1, confidence));
 }
 
-function buildEpisodicEntry(ep, confidence, score, includeProvenance) {
+function buildEpisodicEntry(ep, confidence, score, includeProvenance, contextMatch, moodCongruence) {
   const entry = {
     id: ep.id,
     content: ep.content,
@@ -65,6 +76,12 @@ function buildEpisodicEntry(ep, confidence, score, includeProvenance) {
     source: ep.source,
     createdAt: ep.created_at,
   };
+  if (contextMatch !== undefined) {
+    entry.contextMatch = contextMatch;
+  }
+  if (moodCongruence !== undefined) {
+    entry.moodCongruence = moodCongruence;
+  }
   if (includeProvenance) {
     entry.provenance = {
       source: ep.source,
@@ -121,7 +138,19 @@ function buildProceduralEntry(proc, confidence, score, includeProvenance) {
   return entry;
 }
 
-function knnEpisodic(db, queryBuffer, candidateK, now, minConfidence, includeProvenance, confidenceConfig) {
+function stateClause(includeDormant) {
+  return includeDormant
+    ? "AND (v.state = 'active' OR v.state = 'context_dependent' OR v.state = 'dormant')"
+    : "AND (v.state = 'active' OR v.state = 'context_dependent')";
+}
+
+function matchesDateFilters(createdAt, filters) {
+  if (filters.after && createdAt <= filters.after) return false;
+  if (filters.before && createdAt >= filters.before) return false;
+  return true;
+}
+
+function knnEpisodic(db, queryBuffer, candidateK, now, minConfidence, includeProvenance, confidenceConfig, filters = {}) {
   const rows = db.prepare(`
     SELECT e.*, (1.0 - v.distance) AS similarity
     FROM vec_episodes v
@@ -133,34 +162,51 @@ function knnEpisodic(db, queryBuffer, candidateK, now, minConfidence, includePro
 
   const results = [];
   for (const row of rows) {
-    const confidence = computeEpisodicConfidence(row, now, confidenceConfig);
+    if (!matchesDateFilters(row.created_at, filters)) continue;
+    if (filters.tags?.length) {
+      const rowTags = safeJsonParse(row.tags, []);
+      if (!filters.tags.some(t => rowTags.includes(t))) continue;
+    }
+    if (filters.sources?.length && !filters.sources.includes(row.source)) continue;
+    let confidence = computeEpisodicConfidence(row, now, confidenceConfig);
+
+    let ctxMatch;
+    if (confidenceConfig?.retrievalContext) {
+      const encodingCtx = safeJsonParse(row.context, {});
+      ctxMatch = contextMatchRatio(encodingCtx, confidenceConfig.retrievalContext);
+      confidence *= contextModifier(encodingCtx, confidenceConfig.retrievalContext, confidenceConfig.contextWeight);
+      confidence = Math.max(0, Math.min(1, confidence));
+    }
+
+    let moodMatch;
+    if (confidenceConfig?.retrievalMood) {
+      const encodingAffect = safeJsonParse(row.affect, {});
+      moodMatch = affectSimilarity(encodingAffect, confidenceConfig.retrievalMood);
+      confidence *= moodCongruenceModifier(encodingAffect, confidenceConfig.retrievalMood, confidenceConfig.affectWeight);
+      confidence = Math.max(0, Math.min(1, confidence));
+    }
+
     if (confidence < minConfidence) continue;
     const score = row.similarity * confidence;
-    results.push(buildEpisodicEntry(row, confidence, score, includeProvenance));
+    results.push(buildEpisodicEntry(row, confidence, score, includeProvenance, ctxMatch, moodMatch));
   }
   return results;
 }
 
-function knnSemantic(db, queryBuffer, candidateK, now, minConfidence, includeProvenance, includeDormant, confidenceConfig) {
-  let stateFilter;
-  if (includeDormant) {
-    stateFilter = "AND (v.state = 'active' OR v.state = 'context_dependent' OR v.state = 'dormant')";
-  } else {
-    stateFilter = "AND (v.state = 'active' OR v.state = 'context_dependent')";
-  }
-
+function knnSemantic(db, queryBuffer, candidateK, now, minConfidence, includeProvenance, includeDormant, confidenceConfig, filters = {}) {
   const rows = db.prepare(`
     SELECT s.*, (1.0 - v.distance) AS similarity
     FROM vec_semantics v
     JOIN semantics s ON s.id = v.id
     WHERE v.embedding MATCH ?
       AND k = ?
-      ${stateFilter}
+      ${stateClause(includeDormant)}
   `).all(queryBuffer, candidateK);
 
   const results = [];
   const matchedIds = [];
   for (const row of rows) {
+    if (!matchesDateFilters(row.created_at, filters)) continue;
     const confidence = computeSemanticConfidence(row, now, confidenceConfig);
     if (confidence < minConfidence) continue;
     const score = row.similarity * confidence;
@@ -170,26 +216,20 @@ function knnSemantic(db, queryBuffer, candidateK, now, minConfidence, includePro
   return { results, matchedIds };
 }
 
-function knnProcedural(db, queryBuffer, candidateK, now, minConfidence, includeProvenance, includeDormant, confidenceConfig) {
-  let stateFilter;
-  if (includeDormant) {
-    stateFilter = "AND (v.state = 'active' OR v.state = 'context_dependent' OR v.state = 'dormant')";
-  } else {
-    stateFilter = "AND (v.state = 'active' OR v.state = 'context_dependent')";
-  }
-
+function knnProcedural(db, queryBuffer, candidateK, now, minConfidence, includeProvenance, includeDormant, confidenceConfig, filters = {}) {
   const rows = db.prepare(`
     SELECT p.*, (1.0 - v.distance) AS similarity
     FROM vec_procedures v
     JOIN procedures p ON p.id = v.id
     WHERE v.embedding MATCH ?
       AND k = ?
-      ${stateFilter}
+      ${stateClause(includeDormant)}
   `).all(queryBuffer, candidateK);
 
   const results = [];
   const matchedIds = [];
   for (const row of rows) {
+    if (!matchesDateFilters(row.created_at, filters)) continue;
     const confidence = computeProceduralConfidence(row, now, confidenceConfig);
     if (confidence < minConfidence) continue;
     const score = row.similarity * confidence;
@@ -203,7 +243,7 @@ function knnProcedural(db, queryBuffer, candidateK, now, minConfidence, includeP
  * @param {import('better-sqlite3').Database} db
  * @param {import('./embedding.js').EmbeddingProvider} embeddingProvider
  * @param {string} query
- * @param {{ minConfidence?: number, types?: string[], limit?: number, includeProvenance?: boolean, includeDormant?: boolean }} [options]
+ * @param {{ minConfidence?: number, types?: string[], limit?: number, includeProvenance?: boolean, includeDormant?: boolean, tags?: string[], sources?: string[], after?: string, before?: string }} [options]
  * @returns {AsyncGenerator<{ id: string, content: string, type: string, confidence: number, score: number, source: string, createdAt: string }>}
  */
 export async function* recallStream(db, embeddingProvider, query, options = {}) {
@@ -214,24 +254,30 @@ export async function* recallStream(db, embeddingProvider, query, options = {})
     includeProvenance = false,
     includeDormant = false,
     confidenceConfig,
+    tags,
+    sources,
+    after,
+    before,
   } = options;
 
   const queryVector = await embeddingProvider.embed(query);
   const queryBuffer = embeddingProvider.vectorToBuffer(queryVector);
   const searchTypes = types || ['episodic', 'semantic', 'procedural'];
   const now = new Date();
-  const candidateK = limit * 3;
+  const hasFilters = tags?.length || sources?.length || after || before;
+  const candidateK = hasFilters ? limit * 5 : limit * 3;
+  const filters = { tags, sources, after, before };
 
   const allResults = [];
 
   if (searchTypes.includes('episodic')) {
-    const episodic = knnEpisodic(db, queryBuffer, candidateK, now, minConfidence, includeProvenance, confidenceConfig);
+    const episodic = knnEpisodic(db, queryBuffer, candidateK, now, minConfidence, includeProvenance, confidenceConfig, filters);
     allResults.push(...episodic);
   }
 
   if (searchTypes.includes('semantic')) {
     const { results: semResults, matchedIds: semIds } =
-      knnSemantic(db, queryBuffer, candidateK, now, minConfidence, includeProvenance, includeDormant, confidenceConfig);
+      knnSemantic(db, queryBuffer, candidateK, now, minConfidence, includeProvenance, includeDormant, confidenceConfig, filters);
     allResults.push(...semResults);
 
     if (semIds.length > 0) {
@@ -247,7 +293,7 @@ export async function* recallStream(db, embeddingProvider, query, options = {})
 
   if (searchTypes.includes('procedural')) {
     const { results: procResults, matchedIds: procIds } =
-      knnProcedural(db, queryBuffer, candidateK, now, minConfidence, includeProvenance, includeDormant, confidenceConfig);
+      knnProcedural(db, queryBuffer, candidateK, now, minConfidence, includeProvenance, includeDormant, confidenceConfig, filters);
     allResults.push(...procResults);
 
     if (procIds.length > 0) {
@@ -272,7 +318,7 @@ export async function* recallStream(db, embeddingProvider, query, options = {})
  * @param {import('better-sqlite3').Database} db
  * @param {import('./embedding.js').EmbeddingProvider} embeddingProvider
  * @param {string} query
- * @param {{ minConfidence?: number, types?: string[], limit?: number, includeProvenance?: boolean, includeDormant?: boolean }} [options]
+ * @param {{ minConfidence?: number, types?: string[], limit?: number, includeProvenance?: boolean, includeDormant?: boolean, tags?: string[], sources?: string[], after?: string, before?: string }} [options]
  * @returns {Promise<Array<{ id: string, content: string, type: string, confidence: number, score: number, source: string, createdAt: string }>>}
  */
 export async function recall(db, embeddingProvider, query, options = {}) {