@retrivora-ai/rag-engine 0.2.8 → 0.2.9

package/dist/{chunk-CYVPACA7.mjs → chunk-5U2DHPIX.mjs}

@@ -1556,7 +1556,11 @@ var EmbeddingStrategyResolver = class {
 function normalizeHintValue(value) {
   return value.replace(/\s+/g, " ").trim();
 }
+function isLikelyPromptPhrase(value) {
+  return /^(what|which|who|where|when|why|how)\b/i.test(value.trim());
+}
 function extractQueryFieldHints(question) {
+  var _a, _b, _c;
   if (!question.trim()) return [];
   const hints = /* @__PURE__ */ new Map();
   const addHint = (value, field) => {
@@ -1573,6 +1577,17 @@ function extractQueryFieldHints(question) {
   for (const match of question.matchAll(/["']([^"']{2,100})["']/g)) {
     addHint(match[1]);
   }
+  const naturalQuestionPatterns = [
+    /\b(?:what|which)\s+(?:is|are|was|were)\s+(?:the\s+)?([^?.!,]{1,60}?)\s+of\s+["']?([^"'\n?.!,]{2,120})["']?(?=[?.!,]|$)/gi,
+    /\b(?:who|what)\s+(?:is|are|was|were)\s+["']?([^"'\n?.!,]{2,120})["']?(?=[?.!,]|$)/gi,
+    /\b(?:about|for|regarding)\s+["']?([^"'\n?.!,]{2,120})["']?(?=[?.!,]|$)/gi
+  ];
+  for (const pattern of naturalQuestionPatterns) {
+    for (const match of question.matchAll(pattern)) {
+      const value = (_a = match[2]) != null ? _a : match[1];
+      if (value) addHint(value);
+    }
+  }
   const fieldPattern = `([^\\n:=?.!,]{1,60}?)`;
   const valuePattern = `([^\\n?.!,]{1,120}?)`;
   const fieldValuePatterns = [
@@ -1582,11 +1597,37 @@ function extractQueryFieldHints(question) {
   ];
   for (const pattern of fieldValuePatterns) {
     for (const match of question.matchAll(pattern)) {
-      addHint(match[2], match[1]);
+      const field = normalizeHintValue((_b = match[1]) != null ? _b : "");
+      const value = (_c = match[2]) != null ? _c : "";
+      if (field && !isLikelyPromptPhrase(field)) {
+        addHint(value, field);
+      } else {
+        addHint(value);
+      }
     }
   }
+  for (const match of question.matchAll(/\b[A-Z][a-z]+(?:\s+[A-Z][a-z]+){1,3}\b/g)) {
+    addHint(match[0]);
+  }
   return [...hints.values()];
 }
+function buildQueryFilter(question, hints) {
+  const filter = { metadata: {}, keywords: [], queryText: question };
+  for (const hint of hints) {
+    if (hint.field) {
+      filter.metadata[hint.field] = hint.value;
+    } else {
+      filter.keywords.push(hint.value);
+    }
+  }
+  for (const match of question.matchAll(/\b[A-Z][a-z]+(?:\s+[A-Z][a-z]+){0,3}\b/g)) {
+    const term = normalizeHintValue(match[0]);
+    if (term && !filter.keywords.includes(term)) filter.keywords.push(term);
+  }
+  if (Object.keys(filter.metadata || {}).length === 0) delete filter.metadata;
+  if (filter.keywords && filter.keywords.length === 0) delete filter.keywords;
+  return filter;
+}
 var Pipeline = class {
   constructor(config) {
     this.initialised = false;
@@ -1675,10 +1716,9 @@ var Pipeline = class {
     try {
       const queryVector = await this.embeddingProvider.embed(question, { taskType: "query" });
       const fieldHints = extractQueryFieldHints(question);
-      const rawMatches = await this.vectorDB.query(queryVector, topK, ns, {
-        __queryText: question,
-        __fieldHints: fieldHints
-      });
+      const filter = buildQueryFilter(question, fieldHints);
+      filter.__fieldHints = fieldHints;
+      const rawMatches = await this.vectorDB.query(queryVector, topK, ns, filter);
       const sources = rawMatches.filter((m) => m.score >= scoreThreshold);
       const context = sources.length ? sources.map((m, i) => `[Source ${i + 1}]
 ${m.content}`).join("\n\n---\n\n") : "No relevant context found.";

package/dist/handlers/index.js

@@ -2680,7 +2680,11 @@ var EmbeddingStrategyResolver = class {
 function normalizeHintValue(value) {
   return value.replace(/\s+/g, " ").trim();
 }
+function isLikelyPromptPhrase(value) {
+  return /^(what|which|who|where|when|why|how)\b/i.test(value.trim());
+}
 function extractQueryFieldHints(question) {
+  var _a, _b, _c;
   if (!question.trim()) return [];
   const hints = /* @__PURE__ */ new Map();
   const addHint = (value, field) => {
@@ -2697,6 +2701,17 @@ function extractQueryFieldHints(question) {
   for (const match of question.matchAll(/["']([^"']{2,100})["']/g)) {
     addHint(match[1]);
   }
+  const naturalQuestionPatterns = [
+    /\b(?:what|which)\s+(?:is|are|was|were)\s+(?:the\s+)?([^?.!,]{1,60}?)\s+of\s+["']?([^"'\n?.!,]{2,120})["']?(?=[?.!,]|$)/gi,
+    /\b(?:who|what)\s+(?:is|are|was|were)\s+["']?([^"'\n?.!,]{2,120})["']?(?=[?.!,]|$)/gi,
+    /\b(?:about|for|regarding)\s+["']?([^"'\n?.!,]{2,120})["']?(?=[?.!,]|$)/gi
+  ];
+  for (const pattern of naturalQuestionPatterns) {
+    for (const match of question.matchAll(pattern)) {
+      const value = (_a = match[2]) != null ? _a : match[1];
+      if (value) addHint(value);
+    }
+  }
   const fieldPattern = `([^\\n:=?.!,]{1,60}?)`;
   const valuePattern = `([^\\n?.!,]{1,120}?)`;
   const fieldValuePatterns = [
@@ -2706,11 +2721,37 @@ function extractQueryFieldHints(question) {
   ];
   for (const pattern of fieldValuePatterns) {
     for (const match of question.matchAll(pattern)) {
-      addHint(match[2], match[1]);
+      const field = normalizeHintValue((_b = match[1]) != null ? _b : "");
+      const value = (_c = match[2]) != null ? _c : "";
+      if (field && !isLikelyPromptPhrase(field)) {
+        addHint(value, field);
+      } else {
+        addHint(value);
+      }
     }
   }
+  for (const match of question.matchAll(/\b[A-Z][a-z]+(?:\s+[A-Z][a-z]+){1,3}\b/g)) {
+    addHint(match[0]);
+  }
   return [...hints.values()];
 }
+function buildQueryFilter(question, hints) {
+  const filter = { metadata: {}, keywords: [], queryText: question };
+  for (const hint of hints) {
+    if (hint.field) {
+      filter.metadata[hint.field] = hint.value;
+    } else {
+      filter.keywords.push(hint.value);
+    }
+  }
+  for (const match of question.matchAll(/\b[A-Z][a-z]+(?:\s+[A-Z][a-z]+){0,3}\b/g)) {
+    const term = normalizeHintValue(match[0]);
+    if (term && !filter.keywords.includes(term)) filter.keywords.push(term);
+  }
+  if (Object.keys(filter.metadata || {}).length === 0) delete filter.metadata;
+  if (filter.keywords && filter.keywords.length === 0) delete filter.keywords;
+  return filter;
+}
 var Pipeline = class {
   constructor(config) {
     this.initialised = false;
@@ -2799,10 +2840,9 @@ var Pipeline = class {
     try {
       const queryVector = await this.embeddingProvider.embed(question, { taskType: "query" });
       const fieldHints = extractQueryFieldHints(question);
-      const rawMatches = await this.vectorDB.query(queryVector, topK, ns, {
-        __queryText: question,
-        __fieldHints: fieldHints
-      });
+      const filter = buildQueryFilter(question, fieldHints);
+      filter.__fieldHints = fieldHints;
+      const rawMatches = await this.vectorDB.query(queryVector, topK, ns, filter);
       const sources = rawMatches.filter((m) => m.score >= scoreThreshold);
       const context = sources.length ? sources.map((m, i) => `[Source ${i + 1}]
 ${m.content}`).join("\n\n---\n\n") : "No relevant context found.";

package/dist/handlers/index.mjs

@@ -3,7 +3,7 @@ import {
   createHealthHandler,
   createIngestHandler,
   createUploadHandler
-} from "../chunk-CYVPACA7.mjs";
+} from "../chunk-5U2DHPIX.mjs";
 import "../chunk-EDLTMSNY.mjs";
 import "../chunk-FWCSY2DS.mjs";
 export {

package/dist/server.js

@@ -2769,7 +2769,11 @@ var EmbeddingStrategyResolver = class {
 function normalizeHintValue(value) {
   return value.replace(/\s+/g, " ").trim();
 }
+function isLikelyPromptPhrase(value) {
+  return /^(what|which|who|where|when|why|how)\b/i.test(value.trim());
+}
 function extractQueryFieldHints(question) {
+  var _a, _b, _c;
   if (!question.trim()) return [];
   const hints = /* @__PURE__ */ new Map();
   const addHint = (value, field) => {
@@ -2786,6 +2790,17 @@ function extractQueryFieldHints(question) {
   for (const match of question.matchAll(/["']([^"']{2,100})["']/g)) {
     addHint(match[1]);
   }
+  const naturalQuestionPatterns = [
+    /\b(?:what|which)\s+(?:is|are|was|were)\s+(?:the\s+)?([^?.!,]{1,60}?)\s+of\s+["']?([^"'\n?.!,]{2,120})["']?(?=[?.!,]|$)/gi,
+    /\b(?:who|what)\s+(?:is|are|was|were)\s+["']?([^"'\n?.!,]{2,120})["']?(?=[?.!,]|$)/gi,
+    /\b(?:about|for|regarding)\s+["']?([^"'\n?.!,]{2,120})["']?(?=[?.!,]|$)/gi
+  ];
+  for (const pattern of naturalQuestionPatterns) {
+    for (const match of question.matchAll(pattern)) {
+      const value = (_a = match[2]) != null ? _a : match[1];
+      if (value) addHint(value);
+    }
+  }
   const fieldPattern = `([^\\n:=?.!,]{1,60}?)`;
   const valuePattern = `([^\\n?.!,]{1,120}?)`;
   const fieldValuePatterns = [
@@ -2795,11 +2810,37 @@ function extractQueryFieldHints(question) {
   ];
   for (const pattern of fieldValuePatterns) {
     for (const match of question.matchAll(pattern)) {
-      addHint(match[2], match[1]);
+      const field = normalizeHintValue((_b = match[1]) != null ? _b : "");
+      const value = (_c = match[2]) != null ? _c : "";
+      if (field && !isLikelyPromptPhrase(field)) {
+        addHint(value, field);
+      } else {
+        addHint(value);
+      }
     }
   }
+  for (const match of question.matchAll(/\b[A-Z][a-z]+(?:\s+[A-Z][a-z]+){1,3}\b/g)) {
+    addHint(match[0]);
+  }
   return [...hints.values()];
 }
+function buildQueryFilter(question, hints) {
+  const filter = { metadata: {}, keywords: [], queryText: question };
+  for (const hint of hints) {
+    if (hint.field) {
+      filter.metadata[hint.field] = hint.value;
+    } else {
+      filter.keywords.push(hint.value);
+    }
+  }
+  for (const match of question.matchAll(/\b[A-Z][a-z]+(?:\s+[A-Z][a-z]+){0,3}\b/g)) {
+    const term = normalizeHintValue(match[0]);
+    if (term && !filter.keywords.includes(term)) filter.keywords.push(term);
+  }
+  if (Object.keys(filter.metadata || {}).length === 0) delete filter.metadata;
+  if (filter.keywords && filter.keywords.length === 0) delete filter.keywords;
+  return filter;
+}
 var Pipeline = class {
   constructor(config) {
     this.initialised = false;
@@ -2888,10 +2929,9 @@ var Pipeline = class {
     try {
       const queryVector = await this.embeddingProvider.embed(question, { taskType: "query" });
       const fieldHints = extractQueryFieldHints(question);
-      const rawMatches = await this.vectorDB.query(queryVector, topK, ns, {
-        __queryText: question,
-        __fieldHints: fieldHints
-      });
+      const filter = buildQueryFilter(question, fieldHints);
+      filter.__fieldHints = fieldHints;
+      const rawMatches = await this.vectorDB.query(queryVector, topK, ns, filter);
       const sources = rawMatches.filter((m) => m.score >= scoreThreshold);
       const context = sources.length ? sources.map((m, i) => `[Source ${i + 1}]
 ${m.content}`).join("\n\n---\n\n") : "No relevant context found.";

package/dist/server.mjs

@@ -34,7 +34,7 @@ import {
   createIngestHandler,
   createUploadHandler,
   getRagConfig
-} from "./chunk-CYVPACA7.mjs";
+} from "./chunk-5U2DHPIX.mjs";
 import "./chunk-EDLTMSNY.mjs";
 import {
   PineconeProvider

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@retrivora-ai/rag-engine",
-  "version": "0.2.8",
+  "version": "0.2.9",
   "description": "Retrivora AI is a plug-and-play AI engine for RAG chat experiences — generic vector DB + LLM provider, embeddable or standalone.",
   "author": "Abhinav Alkuchi",
   "license": "MIT",

package/src/core/Pipeline.ts

@@ -12,10 +12,20 @@ interface QueryFieldHint {
   value: string;
 }
 
+interface QueryFilter {
+  metadata?: Record<string, string>;
+  keywords?: string[];
+  queryText?: string;
+}
+
 function normalizeHintValue(value: string): string {
   return value.replace(/\s+/g, ' ').trim();
 }
 
+function isLikelyPromptPhrase(value: string): boolean {
+  return /^(what|which|who|where|when|why|how)\b/i.test(value.trim());
+}
+
 function extractQueryFieldHints(question: string): QueryFieldHint[] {
   if (!question.trim()) return [];
 
@@ -45,6 +55,19 @@ function extractQueryFieldHints(question: string): QueryFieldHint[] {
     addHint(match[1]);
   }
 
+  const naturalQuestionPatterns = [
+    /\b(?:what|which)\s+(?:is|are|was|were)\s+(?:the\s+)?([^?.!,]{1,60}?)\s+of\s+["']?([^"'\n?.!,]{2,120})["']?(?=[?.!,]|$)/gi,
+    /\b(?:who|what)\s+(?:is|are|was|were)\s+["']?([^"'\n?.!,]{2,120})["']?(?=[?.!,]|$)/gi,
+    /\b(?:about|for|regarding)\s+["']?([^"'\n?.!,]{2,120})["']?(?=[?.!,]|$)/gi,
+  ];
+
+  for (const pattern of naturalQuestionPatterns) {
+    for (const match of question.matchAll(pattern)) {
+      const value = match[2] ?? match[1];
+      if (value) addHint(value);
+    }
+  }
+
   const fieldPattern = `([^\\n:=?.!,]{1,60}?)`;
   const valuePattern = `([^\\n?.!,]{1,120}?)`;
   const fieldValuePatterns = [
@@ -55,13 +78,50 @@ function extractQueryFieldHints(question: string): QueryFieldHint[] {
 
   for (const pattern of fieldValuePatterns) {
     for (const match of question.matchAll(pattern)) {
-      addHint(match[2], match[1]);
+      const field = normalizeHintValue(match[1] ?? '');
+      const value = match[2] ?? '';
+
+      if (field && !isLikelyPromptPhrase(field)) {
+        addHint(value, field);
+      } else {
+        addHint(value);
+      }
     }
   }
 
+  for (const match of question.matchAll(/\b[A-Z][a-z]+(?:\s+[A-Z][a-z]+){1,3}\b/g)) {
+    addHint(match[0]);
+  }
+
   return [...hints.values()];
 }
 
+function buildQueryFilter(question: string, hints: QueryFieldHint[]): QueryFilter {
+  const filter: QueryFilter = { metadata: {}, keywords: [], queryText: question };
+
+  for (const hint of hints) {
+    if (hint.field) {
+      // prefer last-seen value for a field; providers may interpret filters differently
+      filter.metadata![hint.field] = hint.value;
+    } else {
+      // treat as keyword / named entity
+      filter.keywords!.push(hint.value);
+    }
+  }
+
+  // Also extract quoted phrases (if not already captured) as keywords
+  for (const match of question.matchAll(/\b[A-Z][a-z]+(?:\s+[A-Z][a-z]+){0,3}\b/g)) {
+    const term = normalizeHintValue(match[0]);
+    if (term && !filter.keywords!.includes(term)) filter.keywords!.push(term);
+  }
+
+  // Remove empty metadata object when no fields present to keep filter minimal
+  if (Object.keys(filter.metadata || {}).length === 0) delete filter.metadata;
+  if (filter.keywords && filter.keywords.length === 0) delete filter.keywords;
+
+  return filter;
+}
+
 /**
  * Pipeline — orchestrates the RAG flow: Embed → Search → Augment → Generate.
  *
@@ -191,11 +251,15 @@ export class Pipeline {
     try {
       const queryVector = await this.embeddingProvider.embed(question, { taskType: 'query' });
       const fieldHints = extractQueryFieldHints(question);
-      // Pass the original text to the vector DB so it can perform hybrid (keyword) search if supported
-      const rawMatches = await this.vectorDB.query(queryVector, topK, ns, {
-        __queryText: question,
-        __fieldHints: fieldHints,
-      });
+      // Build a provider-agnostic filter from hints + original query text. Providers
+      // may choose to use `queryText` for hybrid keyword matching and `metadata`
+      // for exact/filtered metadata queries.
+      const filter = buildQueryFilter(question, fieldHints) as Record<string, unknown>;
+
+      // Keep the raw hints available as well for providers that want richer parsing
+      filter.__fieldHints = fieldHints;
+
+      const rawMatches = await this.vectorDB.query(queryVector, topK, ns, filter);
 
       const sources = rawMatches.filter((m) => m.score >= scoreThreshold);
       const context = sources.length