@retrivora-ai/rag-engine 0.2.2 → 0.2.4

package/dist/{DocumentChunker-BEyzadsv.d.mts → DocumentChunker-BICIjSuG.d.mts}

@@ -18,6 +18,8 @@ interface ChatOptions {
 interface EmbedOptions {
     /** Override model for this specific embed call */
     model?: string;
+    /** Specify the task type for models that require prefixes (e.g. nomic-embed-text) */
+    taskType?: 'query' | 'document';
 }
 interface ILLMProvider {
     /**

package/dist/{DocumentChunker-BEyzadsv.d.ts → DocumentChunker-BICIjSuG.d.ts}

@@ -18,6 +18,8 @@ interface ChatOptions {
 interface EmbedOptions {
     /** Override model for this specific embed call */
     model?: string;
+    /** Specify the task type for models that require prefixes (e.g. nomic-embed-text) */
+    taskType?: 'query' | 'document';
 }
 interface ILLMProvider {
     /**

package/dist/{RagConfig-D_rSf8ep.d.mts → RagConfig-Ttch1N4d.d.mts}

@@ -85,6 +85,10 @@ interface EmbeddingConfig {
     baseUrl?: string;
     /** Output vector dimension — must match the index dimension */
     dimensions?: number;
+    /** Optional prefix to prepend to queries (for models like nomic-embed-text) */
+    queryPrefix?: string;
+    /** Optional prefix to prepend to documents during ingestion */
+    documentPrefix?: string;
     /** Provider-specific options for custom adapters */
     options?: Record<string, unknown>;
 }

package/dist/{RagConfig-D_rSf8ep.d.ts → RagConfig-Ttch1N4d.d.ts}

@@ -85,6 +85,10 @@ interface EmbeddingConfig {
     baseUrl?: string;
     /** Output vector dimension — must match the index dimension */
     dimensions?: number;
+    /** Optional prefix to prepend to queries (for models like nomic-embed-text) */
+    queryPrefix?: string;
+    /** Optional prefix to prepend to documents during ingestion */
+    documentPrefix?: string;
     /** Provider-specific options for custom adapters */
     options?: Record<string, unknown>;
 }

package/dist/{chunk-2VR5ZMXV.mjs → chunk-7K4KXB6G.mjs}

@@ -130,6 +130,8 @@ function getRagConfig(env = process.env) {
       apiKey: embeddingApiKeyByProvider[embeddingProvider],
       baseUrl: readString(env, "EMBEDDING_BASE_URL"),
       dimensions: embeddingDimensions,
+      queryPrefix: readString(env, "EMBEDDING_QUERY_PREFIX"),
+      documentPrefix: readString(env, "EMBEDDING_DOCUMENT_PREFIX"),
       options: {
         profile: readString(env, "EMBEDDING_UNIVERSAL_PROFILE")
       }
@@ -903,13 +905,25 @@ ${context}`;
     return data.message.content;
   }
   async embed(text, options) {
-    var _a, _b, _c, _d, _e, _f, _g;
+    var _a, _b, _c, _d, _e, _f, _g, _h, _i;
     const model = (_c = (_b = options == null ? void 0 : options.model) != null ? _b : (_a = this.embeddingConfig) == null ? void 0 : _a.model) != null ? _c : "nomic-embed-text";
     const baseURL = (_f = (_e = (_d = this.embeddingConfig) == null ? void 0 : _d.baseUrl) != null ? _e : this.llmConfig.baseUrl) != null ? _f : "http://localhost:11434";
     const client = baseURL !== ((_g = this.llmConfig.baseUrl) != null ? _g : "http://localhost:11434") ? axios.create({ baseURL, timeout: 6e4 }) : this.http;
+    let prompt = text;
+    const queryPrefix = (_h = this.embeddingConfig) == null ? void 0 : _h.queryPrefix;
+    const docPrefix = (_i = this.embeddingConfig) == null ? void 0 : _i.documentPrefix;
+    if ((options == null ? void 0 : options.taskType) === "query" && queryPrefix) {
+      if (!prompt.startsWith(queryPrefix)) {
+        prompt = `${queryPrefix}${text}`;
+      }
+    } else if ((options == null ? void 0 : options.taskType) === "document" && docPrefix) {
+      if (!prompt.startsWith(docPrefix)) {
+        prompt = `${docPrefix}${text}`;
+      }
+    }
     const { data } = await client.post("/api/embeddings", {
       model,
-      prompt: text
+      prompt
     });
     return data.embedding;
   }
@@ -1582,7 +1596,7 @@ var Pipeline = class {
         };
         const vectors = await BatchProcessor.mapWithRetry(
           chunks.map((c) => c.content),
-          (text) => this.embeddingProvider.embed(text),
+          (text) => this.embeddingProvider.embed(text, { taskType: "document" }),
           embedBatchOptions
         );
         if (vectors.length !== chunks.length) {
@@ -1625,8 +1639,8 @@ var Pipeline = class {
     const topK = (_b = (_a = this.config.rag) == null ? void 0 : _a.topK) != null ? _b : 5;
     const scoreThreshold = (_d = (_c = this.config.rag) == null ? void 0 : _c.scoreThreshold) != null ? _d : 0;
     try {
-      const queryVector = await this.embeddingProvider.embed(question);
-      const rawMatches = await this.vectorDB.query(queryVector, topK, ns);
+      const queryVector = await this.embeddingProvider.embed(question, { taskType: "query" });
+      const rawMatches = await this.vectorDB.query(queryVector, topK, ns, { __queryText: question });
       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.d.mts

@@ -1,3 +1,3 @@
-import '../RagConfig-D_rSf8ep.mjs';
-export { c as createChatHandler, a as createHealthHandler, b as createIngestHandler, d as createUploadHandler } from '../index-BJ8CUArE.mjs';
+import '../RagConfig-Ttch1N4d.mjs';
+export { c as createChatHandler, a as createHealthHandler, b as createIngestHandler, d as createUploadHandler } from '../index-sbCtrIRT.mjs';
 import 'next/server';

package/dist/handlers/index.d.ts

@@ -1,3 +1,3 @@
-import '../RagConfig-D_rSf8ep.js';
-export { c as createChatHandler, a as createHealthHandler, b as createIngestHandler, d as createUploadHandler } from '../index-DtNprGGj.js';
+import '../RagConfig-Ttch1N4d.js';
+export { c as createChatHandler, a as createHealthHandler, b as createIngestHandler, d as createUploadHandler } from '../index-rK0KAr2S.js';
 import 'next/server';

package/dist/handlers/index.js

@@ -1284,6 +1284,8 @@ function getRagConfig(env = process.env) {
       apiKey: embeddingApiKeyByProvider[embeddingProvider],
       baseUrl: readString(env, "EMBEDDING_BASE_URL"),
       dimensions: embeddingDimensions,
+      queryPrefix: readString(env, "EMBEDDING_QUERY_PREFIX"),
+      documentPrefix: readString(env, "EMBEDDING_DOCUMENT_PREFIX"),
       options: {
         profile: readString(env, "EMBEDDING_UNIVERSAL_PROFILE")
       }
@@ -2057,13 +2059,25 @@ ${context}`;
     return data.message.content;
   }
   async embed(text, options) {
-    var _a, _b, _c, _d, _e, _f, _g;
+    var _a, _b, _c, _d, _e, _f, _g, _h, _i;
     const model = (_c = (_b = options == null ? void 0 : options.model) != null ? _b : (_a = this.embeddingConfig) == null ? void 0 : _a.model) != null ? _c : "nomic-embed-text";
     const baseURL = (_f = (_e = (_d = this.embeddingConfig) == null ? void 0 : _d.baseUrl) != null ? _e : this.llmConfig.baseUrl) != null ? _f : "http://localhost:11434";
     const client = baseURL !== ((_g = this.llmConfig.baseUrl) != null ? _g : "http://localhost:11434") ? import_axios.default.create({ baseURL, timeout: 6e4 }) : this.http;
+    let prompt = text;
+    const queryPrefix = (_h = this.embeddingConfig) == null ? void 0 : _h.queryPrefix;
+    const docPrefix = (_i = this.embeddingConfig) == null ? void 0 : _i.documentPrefix;
+    if ((options == null ? void 0 : options.taskType) === "query" && queryPrefix) {
+      if (!prompt.startsWith(queryPrefix)) {
+        prompt = `${queryPrefix}${text}`;
+      }
+    } else if ((options == null ? void 0 : options.taskType) === "document" && docPrefix) {
+      if (!prompt.startsWith(docPrefix)) {
+        prompt = `${docPrefix}${text}`;
+      }
+    }
     const { data } = await client.post("/api/embeddings", {
       model,
-      prompt: text
+      prompt
     });
     return data.embedding;
   }
@@ -2692,7 +2706,7 @@ var Pipeline = class {
         };
         const vectors = await BatchProcessor.mapWithRetry(
           chunks.map((c) => c.content),
-          (text) => this.embeddingProvider.embed(text),
+          (text) => this.embeddingProvider.embed(text, { taskType: "document" }),
           embedBatchOptions
         );
         if (vectors.length !== chunks.length) {
@@ -2735,8 +2749,8 @@ var Pipeline = class {
     const topK = (_b = (_a = this.config.rag) == null ? void 0 : _a.topK) != null ? _b : 5;
     const scoreThreshold = (_d = (_c = this.config.rag) == null ? void 0 : _c.scoreThreshold) != null ? _d : 0;
     try {
-      const queryVector = await this.embeddingProvider.embed(question);
-      const rawMatches = await this.vectorDB.query(queryVector, topK, ns);
+      const queryVector = await this.embeddingProvider.embed(question, { taskType: "query" });
+      const rawMatches = await this.vectorDB.query(queryVector, topK, ns, { __queryText: question });
       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-2VR5ZMXV.mjs";
+} from "../chunk-7K4KXB6G.mjs";
 import "../chunk-EDLTMSNY.mjs";
 import "../chunk-FWCSY2DS.mjs";
 export {

package/dist/{index-DtNprGGj.d.ts → index-rK0KAr2S.d.ts}

@@ -1,4 +1,4 @@
-import { c as VectorDBConfig, L as LLMConfig, E as EmbeddingConfig, a as RagConfig, C as ChatResponse } from './RagConfig-D_rSf8ep.js';
+import { c as VectorDBConfig, L as LLMConfig, E as EmbeddingConfig, a as RagConfig, C as ChatResponse } from './RagConfig-Ttch1N4d.js';
 import { NextRequest, NextResponse } from 'next/server';
 
 /**

package/dist/{index-BJ8CUArE.d.mts → index-sbCtrIRT.d.mts}

@@ -1,4 +1,4 @@
-import { c as VectorDBConfig, L as LLMConfig, E as EmbeddingConfig, a as RagConfig, C as ChatResponse } from './RagConfig-D_rSf8ep.mjs';
+import { c as VectorDBConfig, L as LLMConfig, E as EmbeddingConfig, a as RagConfig, C as ChatResponse } from './RagConfig-Ttch1N4d.mjs';
 import { NextRequest, NextResponse } from 'next/server';
 
 /**

package/dist/index.d.mts

@@ -1,9 +1,9 @@
 import * as react_jsx_runtime from 'react/jsx-runtime';
 import React$1, { ReactNode } from 'react';
-import { C as ChatMessage } from './DocumentChunker-BEyzadsv.mjs';
-export { a as ChatOptions, b as Chunk, c as ChunkOptions, E as EmbedOptions, I as ILLMProvider } from './DocumentChunker-BEyzadsv.mjs';
-import { V as VectorMatch, U as UIConfig } from './RagConfig-D_rSf8ep.mjs';
-export { C as ChatResponse, E as EmbeddingConfig, I as IngestDocument, L as LLMConfig, R as RAGConfig, a as RagConfig, b as UpsertDocument, c as VectorDBConfig } from './RagConfig-D_rSf8ep.mjs';
+import { C as ChatMessage } from './DocumentChunker-BICIjSuG.mjs';
+export { a as ChatOptions, b as Chunk, c as ChunkOptions, E as EmbedOptions, I as ILLMProvider } from './DocumentChunker-BICIjSuG.mjs';
+import { V as VectorMatch, U as UIConfig } from './RagConfig-Ttch1N4d.mjs';
+export { C as ChatResponse, E as EmbeddingConfig, I as IngestDocument, L as LLMConfig, R as RAGConfig, a as RagConfig, b as UpsertDocument, c as VectorDBConfig } from './RagConfig-Ttch1N4d.mjs';
 
 interface ChatWidgetProps {
     /** Position of the floating button. Defaults to bottom-right. */

package/dist/index.d.ts

@@ -1,9 +1,9 @@
 import * as react_jsx_runtime from 'react/jsx-runtime';
 import React$1, { ReactNode } from 'react';
-import { C as ChatMessage } from './DocumentChunker-BEyzadsv.js';
-export { a as ChatOptions, b as Chunk, c as ChunkOptions, E as EmbedOptions, I as ILLMProvider } from './DocumentChunker-BEyzadsv.js';
-import { V as VectorMatch, U as UIConfig } from './RagConfig-D_rSf8ep.js';
-export { C as ChatResponse, E as EmbeddingConfig, I as IngestDocument, L as LLMConfig, R as RAGConfig, a as RagConfig, b as UpsertDocument, c as VectorDBConfig } from './RagConfig-D_rSf8ep.js';
+import { C as ChatMessage } from './DocumentChunker-BICIjSuG.js';
+export { a as ChatOptions, b as Chunk, c as ChunkOptions, E as EmbedOptions, I as ILLMProvider } from './DocumentChunker-BICIjSuG.js';
+import { V as VectorMatch, U as UIConfig } from './RagConfig-Ttch1N4d.js';
+export { C as ChatResponse, E as EmbeddingConfig, I as IngestDocument, L as LLMConfig, R as RAGConfig, a as RagConfig, b as UpsertDocument, c as VectorDBConfig } from './RagConfig-Ttch1N4d.js';
 
 interface ChatWidgetProps {
     /** Position of the floating button. Defaults to bottom-right. */

package/dist/server.d.mts

@@ -1,9 +1,9 @@
-import { a as RagConfig, C as ChatResponse, I as IngestDocument, c as VectorDBConfig, b as UpsertDocument, V as VectorMatch, L as LLMConfig, E as EmbeddingConfig, d as VectorDBProvider, e as LLMProvider, f as EmbeddingProvider } from './RagConfig-D_rSf8ep.mjs';
-export { R as RAGConfig, U as UIConfig } from './RagConfig-D_rSf8ep.mjs';
-import { C as ChatMessage, I as ILLMProvider, a as ChatOptions, E as EmbedOptions } from './DocumentChunker-BEyzadsv.mjs';
-export { b as Chunk, c as ChunkOptions, D as DocumentChunker } from './DocumentChunker-BEyzadsv.mjs';
-import { H as HealthCheckResult } from './index-BJ8CUArE.mjs';
-export { P as ProviderHealthCheck, c as createChatHandler, a as createHealthHandler, b as createIngestHandler, d as createUploadHandler } from './index-BJ8CUArE.mjs';
+import { a as RagConfig, C as ChatResponse, I as IngestDocument, c as VectorDBConfig, b as UpsertDocument, V as VectorMatch, L as LLMConfig, E as EmbeddingConfig, d as VectorDBProvider, e as LLMProvider, f as EmbeddingProvider } from './RagConfig-Ttch1N4d.mjs';
+export { R as RAGConfig, U as UIConfig } from './RagConfig-Ttch1N4d.mjs';
+import { C as ChatMessage, I as ILLMProvider, a as ChatOptions, E as EmbedOptions } from './DocumentChunker-BICIjSuG.mjs';
+export { b as Chunk, c as ChunkOptions, D as DocumentChunker } from './DocumentChunker-BICIjSuG.mjs';
+import { H as HealthCheckResult } from './index-sbCtrIRT.mjs';
+export { P as ProviderHealthCheck, c as createChatHandler, a as createHealthHandler, b as createIngestHandler, d as createUploadHandler } from './index-sbCtrIRT.mjs';
 import 'next/server';
 
 /**

package/dist/server.d.ts

@@ -1,9 +1,9 @@
-import { a as RagConfig, C as ChatResponse, I as IngestDocument, c as VectorDBConfig, b as UpsertDocument, V as VectorMatch, L as LLMConfig, E as EmbeddingConfig, d as VectorDBProvider, e as LLMProvider, f as EmbeddingProvider } from './RagConfig-D_rSf8ep.js';
-export { R as RAGConfig, U as UIConfig } from './RagConfig-D_rSf8ep.js';
-import { C as ChatMessage, I as ILLMProvider, a as ChatOptions, E as EmbedOptions } from './DocumentChunker-BEyzadsv.js';
-export { b as Chunk, c as ChunkOptions, D as DocumentChunker } from './DocumentChunker-BEyzadsv.js';
-import { H as HealthCheckResult } from './index-DtNprGGj.js';
-export { P as ProviderHealthCheck, c as createChatHandler, a as createHealthHandler, b as createIngestHandler, d as createUploadHandler } from './index-DtNprGGj.js';
+import { a as RagConfig, C as ChatResponse, I as IngestDocument, c as VectorDBConfig, b as UpsertDocument, V as VectorMatch, L as LLMConfig, E as EmbeddingConfig, d as VectorDBProvider, e as LLMProvider, f as EmbeddingProvider } from './RagConfig-Ttch1N4d.js';
+export { R as RAGConfig, U as UIConfig } from './RagConfig-Ttch1N4d.js';
+import { C as ChatMessage, I as ILLMProvider, a as ChatOptions, E as EmbedOptions } from './DocumentChunker-BICIjSuG.js';
+export { b as Chunk, c as ChunkOptions, D as DocumentChunker } from './DocumentChunker-BICIjSuG.js';
+import { H as HealthCheckResult } from './index-rK0KAr2S.js';
+export { P as ProviderHealthCheck, c as createChatHandler, a as createHealthHandler, b as createIngestHandler, d as createUploadHandler } from './index-rK0KAr2S.js';
 import 'next/server';
 
 /**

package/dist/server.js

@@ -1328,6 +1328,8 @@ function getRagConfig(env = process.env) {
       apiKey: embeddingApiKeyByProvider[embeddingProvider],
       baseUrl: readString(env, "EMBEDDING_BASE_URL"),
       dimensions: embeddingDimensions,
+      queryPrefix: readString(env, "EMBEDDING_QUERY_PREFIX"),
+      documentPrefix: readString(env, "EMBEDDING_DOCUMENT_PREFIX"),
       options: {
         profile: readString(env, "EMBEDDING_UNIVERSAL_PROFILE")
       }
@@ -2101,13 +2103,25 @@ ${context}`;
     return data.message.content;
   }
   async embed(text, options) {
-    var _a, _b, _c, _d, _e, _f, _g;
+    var _a, _b, _c, _d, _e, _f, _g, _h, _i;
     const model = (_c = (_b = options == null ? void 0 : options.model) != null ? _b : (_a = this.embeddingConfig) == null ? void 0 : _a.model) != null ? _c : "nomic-embed-text";
     const baseURL = (_f = (_e = (_d = this.embeddingConfig) == null ? void 0 : _d.baseUrl) != null ? _e : this.llmConfig.baseUrl) != null ? _f : "http://localhost:11434";
     const client = baseURL !== ((_g = this.llmConfig.baseUrl) != null ? _g : "http://localhost:11434") ? import_axios.default.create({ baseURL, timeout: 6e4 }) : this.http;
+    let prompt = text;
+    const queryPrefix = (_h = this.embeddingConfig) == null ? void 0 : _h.queryPrefix;
+    const docPrefix = (_i = this.embeddingConfig) == null ? void 0 : _i.documentPrefix;
+    if ((options == null ? void 0 : options.taskType) === "query" && queryPrefix) {
+      if (!prompt.startsWith(queryPrefix)) {
+        prompt = `${queryPrefix}${text}`;
+      }
+    } else if ((options == null ? void 0 : options.taskType) === "document" && docPrefix) {
+      if (!prompt.startsWith(docPrefix)) {
+        prompt = `${docPrefix}${text}`;
+      }
+    }
     const { data } = await client.post("/api/embeddings", {
       model,
-      prompt: text
+      prompt
     });
     return data.embedding;
   }
@@ -2781,7 +2795,7 @@ var Pipeline = class {
         };
         const vectors = await BatchProcessor.mapWithRetry(
           chunks.map((c) => c.content),
-          (text) => this.embeddingProvider.embed(text),
+          (text) => this.embeddingProvider.embed(text, { taskType: "document" }),
           embedBatchOptions
         );
         if (vectors.length !== chunks.length) {
@@ -2824,8 +2838,8 @@ var Pipeline = class {
     const topK = (_b = (_a = this.config.rag) == null ? void 0 : _a.topK) != null ? _b : 5;
     const scoreThreshold = (_d = (_c = this.config.rag) == null ? void 0 : _c.scoreThreshold) != null ? _d : 0;
     try {
-      const queryVector = await this.embeddingProvider.embed(question);
-      const rawMatches = await this.vectorDB.query(queryVector, topK, ns);
+      const queryVector = await this.embeddingProvider.embed(question, { taskType: "query" });
+      const rawMatches = await this.vectorDB.query(queryVector, topK, ns, { __queryText: question });
       const sources = rawMatches.filter((m) => m.score >= scoreThreshold);
       const context = sources.length ? sources.map((m, i) => `[Source ${i + 1}]
 ${m.content}`).join("\n\n---\n\n") : "No relevant context found.";
@@ -3756,29 +3770,46 @@ var MultiTablePostgresProvider = class extends BaseVectorProvider {
     const vectorLiteral = `[${vector.join(",")}]`;
     const allResults = [];
     console.log(`[MultiTablePostgresProvider] --- Starting Multi-Table Search ---`);
+    const queryText = _filter == null ? void 0 : _filter.__queryText;
     for (const table of this.tables) {
       try {
-        const result = await this.pool.query(
-          `SELECT *, 
-           (1 - (embedding <=> $1::vector)) AS vector_score
-           FROM "${table}"
-           ORDER BY vector_score DESC
-           LIMIT 50`,
-          [vectorLiteral]
-        );
+        let sqlQuery = "";
+        let params = [];
+        if (queryText) {
+          sqlQuery = `
+            SELECT *, 
+             (1 - (embedding <=> $1::vector)) AS vector_score,
+             ts_rank(to_tsvector('english', t.*::text), plainto_tsquery('english', $2)) AS keyword_score,
+             ((1 - (embedding <=> $1::vector)) + (ts_rank(to_tsvector('english', t.*::text), plainto_tsquery('english', $2)) * 2.0)) AS hybrid_score
+            FROM "${table}" t
+            ORDER BY hybrid_score DESC
+            LIMIT 50
+          `;
+          params = [vectorLiteral, queryText];
+        } else {
+          sqlQuery = `
+            SELECT *, 
+             (1 - (embedding <=> $1::vector)) AS hybrid_score
+            FROM "${table}" t
+            ORDER BY hybrid_score DESC
+            LIMIT 50
+          `;
+          params = [vectorLiteral];
+        }
+        const result = await this.pool.query(sqlQuery, params);
         if (result.rowCount && result.rowCount > 0) {
-          console.log(`   \u2705 Table "${table}": Found ${result.rowCount} potential matches. Top score: ${result.rows[0].vector_score.toFixed(4)}`);
+          console.log(`   \u2705 Table "${table}": Found ${result.rowCount} potential matches. Top score: ${result.rows[0].hybrid_score.toFixed(4)}`);
         } else {
           console.log(`   \u26AA Table "${table}": No relevant matches found.`);
         }
         for (const row of result.rows) {
-          const _a = row, { vector_score, id } = _a, rest = __objRest(_a, ["vector_score", "id"]);
+          const _a = row, { hybrid_score, vector_score, keyword_score, id } = _a, rest = __objRest(_a, ["hybrid_score", "vector_score", "keyword_score", "id"]);
           delete rest.embedding;
           const content = `[TYPE: ${table.replace(/s$/, "").toUpperCase()}]
 ` + Object.entries(rest).filter(([k, v]) => v !== null && typeof v !== "object" && k !== "id").map(([k, v]) => `${k}: ${v}`).join("\n");
           allResults.push({
             id: `${table}-${id}`,
-            score: parseFloat(String(vector_score)),
+            score: parseFloat(String(hybrid_score)),
             content,
             metadata: __spreadProps(__spreadValues({}, rest), { source_table: table })
           });

package/dist/server.mjs

@@ -34,7 +34,7 @@ import {
   createIngestHandler,
   createUploadHandler,
   getRagConfig
-} from "./chunk-2VR5ZMXV.mjs";
+} from "./chunk-7K4KXB6G.mjs";
 import "./chunk-EDLTMSNY.mjs";
 import {
   PineconeProvider
@@ -413,29 +413,46 @@ var MultiTablePostgresProvider = class extends BaseVectorProvider {
     const vectorLiteral = `[${vector.join(",")}]`;
     const allResults = [];
     console.log(`[MultiTablePostgresProvider] --- Starting Multi-Table Search ---`);
+    const queryText = _filter == null ? void 0 : _filter.__queryText;
     for (const table of this.tables) {
       try {
-        const result = await this.pool.query(
-          `SELECT *, 
-           (1 - (embedding <=> $1::vector)) AS vector_score
-           FROM "${table}"
-           ORDER BY vector_score DESC
-           LIMIT 50`,
-          [vectorLiteral]
-        );
+        let sqlQuery = "";
+        let params = [];
+        if (queryText) {
+          sqlQuery = `
+            SELECT *, 
+             (1 - (embedding <=> $1::vector)) AS vector_score,
+             ts_rank(to_tsvector('english', t.*::text), plainto_tsquery('english', $2)) AS keyword_score,
+             ((1 - (embedding <=> $1::vector)) + (ts_rank(to_tsvector('english', t.*::text), plainto_tsquery('english', $2)) * 2.0)) AS hybrid_score
+            FROM "${table}" t
+            ORDER BY hybrid_score DESC
+            LIMIT 50
+          `;
+          params = [vectorLiteral, queryText];
+        } else {
+          sqlQuery = `
+            SELECT *, 
+             (1 - (embedding <=> $1::vector)) AS hybrid_score
+            FROM "${table}" t
+            ORDER BY hybrid_score DESC
+            LIMIT 50
+          `;
+          params = [vectorLiteral];
+        }
+        const result = await this.pool.query(sqlQuery, params);
         if (result.rowCount && result.rowCount > 0) {
-          console.log(`   \u2705 Table "${table}": Found ${result.rowCount} potential matches. Top score: ${result.rows[0].vector_score.toFixed(4)}`);
+          console.log(`   \u2705 Table "${table}": Found ${result.rowCount} potential matches. Top score: ${result.rows[0].hybrid_score.toFixed(4)}`);
         } else {
           console.log(`   \u26AA Table "${table}": No relevant matches found.`);
         }
         for (const row of result.rows) {
-          const _a = row, { vector_score, id } = _a, rest = __objRest(_a, ["vector_score", "id"]);
+          const _a = row, { hybrid_score, vector_score, keyword_score, id } = _a, rest = __objRest(_a, ["hybrid_score", "vector_score", "keyword_score", "id"]);
           delete rest.embedding;
           const content = `[TYPE: ${table.replace(/s$/, "").toUpperCase()}]
 ` + Object.entries(rest).filter(([k, v]) => v !== null && typeof v !== "object" && k !== "id").map(([k, v]) => `${k}: ${v}`).join("\n");
           allResults.push({
             id: `${table}-${id}`,
-            score: parseFloat(String(vector_score)),
+            score: parseFloat(String(hybrid_score)),
             content,
             metadata: __spreadProps(__spreadValues({}, rest), { source_table: table })
           });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@retrivora-ai/rag-engine",
-  "version": "0.2.2",
+  "version": "0.2.4",
   "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/config/RagConfig.ts

@@ -81,6 +81,10 @@ export interface EmbeddingConfig {
   baseUrl?: string;
   /** Output vector dimension — must match the index dimension */
   dimensions?: number;
+  /** Optional prefix to prepend to queries (for models like nomic-embed-text) */
+  queryPrefix?: string;
+  /** Optional prefix to prepend to documents during ingestion */
+  documentPrefix?: string;
   /** Provider-specific options for custom adapters */
   options?: Record<string, unknown>;
 }

package/src/config/serverConfig.ts

@@ -150,6 +150,8 @@ export function getRagConfig(env: Record<string, string | undefined> = process.e
       apiKey: embeddingApiKeyByProvider[embeddingProvider],
       baseUrl: readString(env, 'EMBEDDING_BASE_URL'),
       dimensions: embeddingDimensions,
+      queryPrefix: readString(env, 'EMBEDDING_QUERY_PREFIX'),
+      documentPrefix: readString(env, 'EMBEDDING_DOCUMENT_PREFIX'),
       options: {
         profile: readString(env, 'EMBEDDING_UNIVERSAL_PROFILE'),
       },

package/src/core/Pipeline.ts

@@ -82,7 +82,7 @@ export class Pipeline {
 
         const vectors = await BatchProcessor.mapWithRetry(
           chunks.map(c => c.content),
-          (text) => this.embeddingProvider.embed(text),
+          (text) => this.embeddingProvider.embed(text, { taskType: 'document' }),
           embedBatchOptions
         );
 
@@ -134,8 +134,9 @@ export class Pipeline {
     const scoreThreshold = this.config.rag?.scoreThreshold ?? 0.0;
 
     try {
-      const queryVector = await this.embeddingProvider.embed(question);
-      const rawMatches = await this.vectorDB.query(queryVector, topK, ns);
+      const queryVector = await this.embeddingProvider.embed(question, { taskType: 'query' });
+      // 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 });
 
       const sources = rawMatches.filter((m) => m.score >= scoreThreshold);
       const context = sources.length

package/src/llm/ILLMProvider.ts

@@ -21,6 +21,8 @@ export interface ChatOptions {
 export interface EmbedOptions {
   /** Override model for this specific embed call */
   model?: string;
+  /** Specify the task type for models that require prefixes (e.g. nomic-embed-text) */
+  taskType?: 'query' | 'document';
 }
 
 export interface ILLMProvider {

package/src/llm/providers/OllamaProvider.ts

@@ -73,9 +73,25 @@ export class OllamaProvider implements ILLMProvider {
         ? axios.create({ baseURL, timeout: 60_000 })
         : this.http;
 
+    let prompt = text;
+    
+    // Dynamically apply prefixes from configuration if they exist
+    const queryPrefix = this.embeddingConfig?.queryPrefix;
+    const docPrefix = this.embeddingConfig?.documentPrefix;
+    
+    if (options?.taskType === 'query' && queryPrefix) {
+      if (!prompt.startsWith(queryPrefix)) {
+        prompt = `${queryPrefix}${text}`;
+      }
+    } else if (options?.taskType === 'document' && docPrefix) {
+      if (!prompt.startsWith(docPrefix)) {
+        prompt = `${docPrefix}${text}`;
+      }
+    }
+
     const { data } = await client.post<OllamaEmbedResponse>('/api/embeddings', {
       model,
-      prompt: text,
+      prompt,
     });
 
     return data.embedding;

package/src/providers/vectordb/MultiTablePostgresProvider.ts

@@ -46,7 +46,7 @@ export class MultiTablePostgresProvider extends BaseVectorProvider {
 
   async initialize(): Promise<void> {
     this.pool = new Pool({ connectionString: this.connectionString });
-    
+
     // Dynamically discover all tables that have an 'embedding' column
     const client: PoolClient = await this.pool.connect();
     try {
@@ -56,9 +56,9 @@ export class MultiTablePostgresProvider extends BaseVectorProvider {
         WHERE column_name = 'embedding' 
         AND table_schema = 'public'
       `);
-      
+
       const discoveredTables = res.rows.map(r => r.table_name);
-      
+
       if (this.tables.length === 0) {
         // No static tables configured, use everything discovered
         this.tables = discoveredTables;
@@ -66,7 +66,7 @@ export class MultiTablePostgresProvider extends BaseVectorProvider {
         // Filter configured tables to ensure they actually exist with embeddings
         const staticTables = [...this.tables];
         this.tables = staticTables.filter(t => discoveredTables.includes(t));
-        
+
         if (this.tables.length < staticTables.length) {
           const missing = staticTables.filter(t => !discoveredTables.includes(t));
           console.warn(`[MultiTablePostgresProvider] Some configured tables were skipped (missing 'embedding' column): ${missing.join(', ')}`);
@@ -123,31 +123,50 @@ export class MultiTablePostgresProvider extends BaseVectorProvider {
 
     console.log(`[MultiTablePostgresProvider] --- Starting Multi-Table Search ---`);
 
+    const queryText = _filter?.__queryText as string | undefined;
+
     for (const table of this.tables) {
       try {
-        // We use a hybrid approach:
-        // 1. Vector similarity (Semantic)
-        // 2. Simple keyword matching (Exact) to boost relevance for specific names
-        const result = await this.pool.query(
-          `SELECT *, 
-           (1 - (embedding <=> $1::vector)) AS vector_score
-           FROM "${table}"
-           ORDER BY vector_score DESC
-           LIMIT 50`,
-          [vectorLiteral]
-        );
+        let sqlQuery = '';
+        let params: any[] = [];
+
+        if (queryText) {
+          // Actual Hybrid Search: Semantic Vector Score + Full-Text Keyword Rank
+          sqlQuery = `
+            SELECT *, 
+             (1 - (embedding <=> $1::vector)) AS vector_score,
+             ts_rank(to_tsvector('english', t.*::text), plainto_tsquery('english', $2)) AS keyword_score,
+             ((1 - (embedding <=> $1::vector)) + (ts_rank(to_tsvector('english', t.*::text), plainto_tsquery('english', $2)) * 2.0)) AS hybrid_score
+            FROM "${table}" t
+            ORDER BY hybrid_score DESC
+            LIMIT 50
+          `;
+          params = [vectorLiteral, queryText];
+        } else {
+          // Fallback to Vector-only Search
+          sqlQuery = `
+            SELECT *, 
+             (1 - (embedding <=> $1::vector)) AS hybrid_score
+            FROM "${table}" t
+            ORDER BY hybrid_score DESC
+            LIMIT 50
+          `;
+          params = [vectorLiteral];
+        }
+
+        const result = await this.pool.query(sqlQuery, params);
 
         if (result.rowCount && result.rowCount > 0) {
-          console.log(`   ✅ Table "${table}": Found ${result.rowCount} potential matches. Top score: ${result.rows[0].vector_score.toFixed(4)}`);
+          console.log(`   ✅ Table "${table}": Found ${result.rowCount} potential matches. Top score: ${result.rows[0].hybrid_score.toFixed(4)}`);
         } else {
           console.log(`   ⚪ Table "${table}": No relevant matches found.`);
         }
 
         for (const row of result.rows) {
-          const { vector_score, id, ...rest } = row as Record<string, unknown>;
+          const { hybrid_score, vector_score, keyword_score, id, ...rest } = row as Record<string, unknown>;
           delete rest.embedding;
 
-          const content = `[TYPE: ${table.replace(/s$/, '').toUpperCase()}]\n` + 
+          const content = `[TYPE: ${table.replace(/s$/, '').toUpperCase()}]\n` +
             Object.entries(rest)
               .filter(([k, v]) => v !== null && typeof v !== 'object' && k !== 'id')
               .map(([k, v]) => `${k}: ${v}`)
@@ -155,7 +174,7 @@ export class MultiTablePostgresProvider extends BaseVectorProvider {
 
           allResults.push({
             id: `${table}-${id}`,
-            score: parseFloat(String(vector_score)),
+            score: parseFloat(String(hybrid_score)),
             content,
             metadata: { ...rest, source_table: table },
           });
@@ -179,7 +198,7 @@ export class MultiTablePostgresProvider extends BaseVectorProvider {
     // 2. Take top 3 from EACH table guaranteed, then fill the rest with overall best
     const balancedResults: VectorMatch[] = [];
     const tables = Object.keys(resultsByTable);
-    
+
     // First, ensure every table's best match is included
     for (const table of tables) {
       balancedResults.push(...resultsByTable[table].slice(0, 3));