@retrivora-ai/rag-engine 0.2.1 → 0.2.3

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-K53N7SEE.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) {
@@ -1625,7 +1639,7 @@ 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 queryVector = await this.embeddingProvider.embed(question, { taskType: "query" });
       const rawMatches = await this.vectorDB.query(queryVector, topK, ns);
       const sources = rawMatches.filter((m) => m.score >= scoreThreshold);
       const context = sources.length ? sources.map((m, i) => `[Source ${i + 1}]

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) {
@@ -2735,7 +2749,7 @@ 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 queryVector = await this.embeddingProvider.embed(question, { taskType: "query" });
       const rawMatches = await this.vectorDB.query(queryVector, topK, ns);
       const sources = rawMatches.filter((m) => m.score >= scoreThreshold);
       const context = sources.length ? sources.map((m, i) => `[Source ${i + 1}]

package/dist/handlers/index.mjs

@@ -3,7 +3,7 @@ import {
   createHealthHandler,
   createIngestHandler,
   createUploadHandler
-} from "../chunk-2VR5ZMXV.mjs";
+} from "../chunk-K53N7SEE.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) {
@@ -2824,7 +2838,7 @@ 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 queryVector = await this.embeddingProvider.embed(question, { taskType: "query" });
       const rawMatches = await this.vectorDB.query(queryVector, topK, ns);
       const sources = rawMatches.filter((m) => m.score >= scoreThreshold);
       const context = sources.length ? sources.map((m, i) => `[Source ${i + 1}]
@@ -3749,31 +3763,36 @@ var MultiTablePostgresProvider = class extends BaseVectorProvider {
    * Query all configured tables and merge results, sorted by cosine similarity score.
    */
   async query(vector, topK, _namespace, _filter) {
-    var _b, _c;
+    var _b, _c, _d;
     if (!this.pool) {
       throw new Error("[MultiTablePostgresProvider] Provider not initialized. Call initialize() first.");
     }
-    console.log(`[MultiTablePostgresProvider] Querying ${this.tables.length} table(s) with vector dim=${vector.length}`);
     const vectorLiteral = `[${vector.join(",")}]`;
     const allResults = [];
+    console.log(`[MultiTablePostgresProvider] --- Starting Multi-Table Search ---`);
     for (const table of this.tables) {
       try {
         const result = await this.pool.query(
-          `SELECT *, 1 - (embedding <=> $1::vector) AS score
+          `SELECT *, 
+           (1 - (embedding <=> $1::vector)) AS vector_score
            FROM "${table}"
-           ORDER BY embedding <=> $1::vector
+           ORDER BY vector_score DESC
            LIMIT 50`,
           [vectorLiteral]
         );
+        if (result.rowCount && result.rowCount > 0) {
+          console.log(`   \u2705 Table "${table}": Found ${result.rowCount} potential matches. Top score: ${result.rows[0].vector_score.toFixed(4)}`);
+        } else {
+          console.log(`   \u26AA Table "${table}": No relevant matches found.`);
+        }
         for (const row of result.rows) {
-          const _a = row, { score, id } = _a, rest = __objRest(_a, ["score", "id"]);
+          const _a = row, { vector_score, id } = _a, rest = __objRest(_a, ["vector_score", "id"]);
           delete rest.embedding;
-          const type = table.replace(/s$/, "").toUpperCase();
-          const content = `[TYPE: ${type}]
+          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}-${id}`,
-            score: parseFloat(String(score)),
+            score: parseFloat(String(vector_score)),
             content,
             metadata: __spreadProps(__spreadValues({}, rest), { source_table: table })
           });
@@ -3785,11 +3804,23 @@ var MultiTablePostgresProvider = class extends BaseVectorProvider {
         );
       }
     }
-    const sortedResults = allResults.sort((a, b) => b.score - a.score);
-    if (sortedResults.length > 0) {
-      console.log(`[MultiTablePostgresProvider] Top match from "${(_c = (_b = sortedResults[0].metadata) == null ? void 0 : _b.source_table) != null ? _c : "unknown"}" with score ${sortedResults[0].score.toFixed(4)}`);
+    const resultsByTable = {};
+    for (const res of allResults) {
+      const table = (_b = res.metadata) == null ? void 0 : _b.source_table;
+      if (!resultsByTable[table]) resultsByTable[table] = [];
+      resultsByTable[table].push(res);
+    }
+    const balancedResults = [];
+    const tables = Object.keys(resultsByTable);
+    for (const table of tables) {
+      balancedResults.push(...resultsByTable[table].slice(0, 3));
+    }
+    const finalSorted = balancedResults.sort((a, b) => b.score - a.score);
+    if (finalSorted.length > 0) {
+      console.log(`[MultiTablePostgresProvider] --- Search Complete ---`);
+      console.log(`[MultiTablePostgresProvider] Final top match from "${(_d = (_c = finalSorted[0].metadata) == null ? void 0 : _c.source_table) != null ? _d : "unknown"}" with score ${finalSorted[0].score.toFixed(4)}`);
     }
-    return sortedResults.slice(0, Math.max(topK, 15));
+    return finalSorted.slice(0, Math.max(topK, 15));
   }
   async delete(_id, _namespace) {
     console.warn("[MultiTablePostgresProvider] delete() is a no-op for multi-table mode.");

package/dist/server.mjs

@@ -34,7 +34,7 @@ import {
   createIngestHandler,
   createUploadHandler,
   getRagConfig
-} from "./chunk-2VR5ZMXV.mjs";
+} from "./chunk-K53N7SEE.mjs";
 import "./chunk-EDLTMSNY.mjs";
 import {
   PineconeProvider
@@ -406,31 +406,36 @@ var MultiTablePostgresProvider = class extends BaseVectorProvider {
    * Query all configured tables and merge results, sorted by cosine similarity score.
    */
   async query(vector, topK, _namespace, _filter) {
-    var _b, _c;
+    var _b, _c, _d;
     if (!this.pool) {
       throw new Error("[MultiTablePostgresProvider] Provider not initialized. Call initialize() first.");
     }
-    console.log(`[MultiTablePostgresProvider] Querying ${this.tables.length} table(s) with vector dim=${vector.length}`);
     const vectorLiteral = `[${vector.join(",")}]`;
     const allResults = [];
+    console.log(`[MultiTablePostgresProvider] --- Starting Multi-Table Search ---`);
     for (const table of this.tables) {
       try {
         const result = await this.pool.query(
-          `SELECT *, 1 - (embedding <=> $1::vector) AS score
+          `SELECT *, 
+           (1 - (embedding <=> $1::vector)) AS vector_score
            FROM "${table}"
-           ORDER BY embedding <=> $1::vector
+           ORDER BY vector_score DESC
            LIMIT 50`,
           [vectorLiteral]
         );
+        if (result.rowCount && result.rowCount > 0) {
+          console.log(`   \u2705 Table "${table}": Found ${result.rowCount} potential matches. Top score: ${result.rows[0].vector_score.toFixed(4)}`);
+        } else {
+          console.log(`   \u26AA Table "${table}": No relevant matches found.`);
+        }
         for (const row of result.rows) {
-          const _a = row, { score, id } = _a, rest = __objRest(_a, ["score", "id"]);
+          const _a = row, { vector_score, id } = _a, rest = __objRest(_a, ["vector_score", "id"]);
           delete rest.embedding;
-          const type = table.replace(/s$/, "").toUpperCase();
-          const content = `[TYPE: ${type}]
+          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}-${id}`,
-            score: parseFloat(String(score)),
+            score: parseFloat(String(vector_score)),
             content,
             metadata: __spreadProps(__spreadValues({}, rest), { source_table: table })
           });
@@ -442,11 +447,23 @@ var MultiTablePostgresProvider = class extends BaseVectorProvider {
         );
       }
     }
-    const sortedResults = allResults.sort((a, b) => b.score - a.score);
-    if (sortedResults.length > 0) {
-      console.log(`[MultiTablePostgresProvider] Top match from "${(_c = (_b = sortedResults[0].metadata) == null ? void 0 : _b.source_table) != null ? _c : "unknown"}" with score ${sortedResults[0].score.toFixed(4)}`);
+    const resultsByTable = {};
+    for (const res of allResults) {
+      const table = (_b = res.metadata) == null ? void 0 : _b.source_table;
+      if (!resultsByTable[table]) resultsByTable[table] = [];
+      resultsByTable[table].push(res);
+    }
+    const balancedResults = [];
+    const tables = Object.keys(resultsByTable);
+    for (const table of tables) {
+      balancedResults.push(...resultsByTable[table].slice(0, 3));
+    }
+    const finalSorted = balancedResults.sort((a, b) => b.score - a.score);
+    if (finalSorted.length > 0) {
+      console.log(`[MultiTablePostgresProvider] --- Search Complete ---`);
+      console.log(`[MultiTablePostgresProvider] Final top match from "${(_d = (_c = finalSorted[0].metadata) == null ? void 0 : _c.source_table) != null ? _d : "unknown"}" with score ${finalSorted[0].score.toFixed(4)}`);
     }
-    return sortedResults.slice(0, Math.max(topK, 15));
+    return finalSorted.slice(0, Math.max(topK, 15));
   }
   async delete(_id, _namespace) {
     console.warn("[MultiTablePostgresProvider] delete() is a no-op for multi-table mode.");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@retrivora-ai/rag-engine",
-  "version": "0.2.1",
+  "version": "0.2.3",
   "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
         );
 
@@ -134,7 +134,7 @@ export class Pipeline {
     const scoreThreshold = this.config.rag?.scoreThreshold ?? 0.0;
 
     try {
-      const queryVector = await this.embeddingProvider.embed(question);
+      const queryVector = await this.embeddingProvider.embed(question, { taskType: 'query' });
       const rawMatches = await this.vectorDB.query(queryVector, topK, ns);
 
       const sources = rawMatches.filter((m) => m.score >= scoreThreshold);

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

@@ -118,30 +118,36 @@ export class MultiTablePostgresProvider extends BaseVectorProvider {
       throw new Error('[MultiTablePostgresProvider] Provider not initialized. Call initialize() first.');
     }
 
-    console.log(`[MultiTablePostgresProvider] Querying ${this.tables.length} table(s) with vector dim=${vector.length}`);
-
     const vectorLiteral = `[${vector.join(',')}]`;
     const allResults: VectorMatch[] = [];
 
+    console.log(`[MultiTablePostgresProvider] --- Starting Multi-Table Search ---`);
+
     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 score
+          `SELECT *, 
+           (1 - (embedding <=> $1::vector)) AS vector_score
            FROM "${table}"
-           ORDER BY embedding <=> $1::vector
+           ORDER BY vector_score DESC
            LIMIT 50`,
           [vectorLiteral]
         );
 
+        if (result.rowCount && result.rowCount > 0) {
+          console.log(`   ✅ Table "${table}": Found ${result.rowCount} potential matches. Top score: ${result.rows[0].vector_score.toFixed(4)}`);
+        } else {
+          console.log(`   ⚪ Table "${table}": No relevant matches found.`);
+        }
+
         for (const row of result.rows) {
-          const { score, id, ...rest } = row as Record<string, unknown>;
-          // Remove embedding from rest to avoid including it in content
+          const { vector_score, id, ...rest } = row as Record<string, unknown>;
           delete rest.embedding;
 
-          // Build a human-readable content string from ALL available columns.
-          // Prepend the entity type (derived from table name) to prevent cross-table confusion.
-          const type = table.replace(/s$/, '').toUpperCase();
-          const content = `[TYPE: ${type}]\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}`)
@@ -149,7 +155,7 @@ export class MultiTablePostgresProvider extends BaseVectorProvider {
 
           allResults.push({
             id: `${table}-${id}`,
-            score: parseFloat(String(score)),
+            score: parseFloat(String(vector_score)),
             content,
             metadata: { ...rest, source_table: table },
           });
@@ -162,14 +168,32 @@ export class MultiTablePostgresProvider extends BaseVectorProvider {
       }
     }
 
-    // Merge and rank by score descending
-    const sortedResults = allResults.sort((a, b) => b.score - a.score);
+    // 1. Group results by table to ensure we take the top ones from each
+    const resultsByTable: Record<string, VectorMatch[]> = {};
+    for (const res of allResults) {
+      const table = res.metadata?.source_table as string;
+      if (!resultsByTable[table]) resultsByTable[table] = [];
+      resultsByTable[table].push(res);
+    }
+
+    // 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));
+    }
+
+    // 3. Sort the balanced list by score
+    const finalSorted = balancedResults.sort((a, b) => b.score - a.score);
 
-    if (sortedResults.length > 0) {
-      console.log(`[MultiTablePostgresProvider] Top match from "${sortedResults[0].metadata?.source_table ?? 'unknown'}" with score ${sortedResults[0].score.toFixed(4)}`);
+    if (finalSorted.length > 0) {
+      console.log(`[MultiTablePostgresProvider] --- Search Complete ---`);
+      console.log(`[MultiTablePostgresProvider] Final top match from "${finalSorted[0].metadata?.source_table ?? 'unknown'}" with score ${finalSorted[0].score.toFixed(4)}`);
     }
 
-    return sortedResults.slice(0, Math.max(topK, 15));
+    return finalSorted.slice(0, Math.max(topK, 15));
   }
 
   async delete(_id: string | number, _namespace?: string): Promise<void> { // eslint-disable-line @typescript-eslint/no-unused-vars