yt-embeddings-strapi-plugin 0.0.1 → 0.1.1

package/dist/_chunks/{App-Cv1cdLAr.js → App-57XBOCeK.js}

@@ -7,7 +7,7 @@ const react = require("react");
 const styled = require("styled-components");
 const designSystem = require("@strapi/design-system");
 const icons = require("@strapi/icons");
-const index = require("./index-BAfBs5PQ.js");
+const index = require("./index-CZeQ6zu2.js");
 const qs = require("qs");
 const ReactMarkdown = require("react-markdown");
 const _interopDefault = (e) => e && e.__esModule ? e : { default: e };
@@ -488,7 +488,15 @@ function VideoDetails() {
       get(`/${index.PLUGIN_ID}/yt/videos/${videoId}`),
       get(`/${index.PLUGIN_ID}/yt/videos/${videoId}/chunks`)
     ]).then(([videoRes, chunksRes]) => {
-      setVideo(videoRes?.data?.data || videoRes?.data || null);
+      const v = videoRes?.data?.data || videoRes?.data || null;
+      if (v && typeof v.key_moments === "string") {
+        try {
+          v.key_moments = JSON.parse(v.key_moments);
+        } catch {
+          v.key_moments = [];
+        }
+      }
+      setVideo(v);
       setChunks(chunksRes?.data?.data || chunksRes?.data || []);
     }).catch((err) => console.error("Failed to load video:", err)).finally(() => setIsLoading(false));
   }, [videoId, get]);

package/dist/_chunks/{App-bN58O1bN.mjs → App-D7Dy3o44.mjs}

@@ -5,7 +5,7 @@ import { useRef, useState, useEffect, useCallback } from "react";
 import styled from "styled-components";
 import { Button, Modal, Box, TextInput, Link, Accordion, Typography, Tr, Main, Flex, Loader, Table, Thead, Th, Tbody, Td, Badge } from "@strapi/design-system";
 import { ArrowClockwise, ArrowLeft } from "@strapi/icons";
-import { R as RobotIcon, P as PLUGIN_ID } from "./index-K6X5FM2O.mjs";
+import { R as RobotIcon, P as PLUGIN_ID } from "./index-CuxGkc_5.mjs";
 import qs from "qs";
 import ReactMarkdown from "react-markdown";
 const MarkdownWrapper = styled.div`
@@ -482,7 +482,15 @@ function VideoDetails() {
       get(`/${PLUGIN_ID}/yt/videos/${videoId}`),
       get(`/${PLUGIN_ID}/yt/videos/${videoId}/chunks`)
     ]).then(([videoRes, chunksRes]) => {
-      setVideo(videoRes?.data?.data || videoRes?.data || null);
+      const v = videoRes?.data?.data || videoRes?.data || null;
+      if (v && typeof v.key_moments === "string") {
+        try {
+          v.key_moments = JSON.parse(v.key_moments);
+        } catch {
+          v.key_moments = [];
+        }
+      }
+      setVideo(v);
       setChunks(chunksRes?.data?.data || chunksRes?.data || []);
     }).catch((err) => console.error("Failed to load video:", err)).finally(() => setIsLoading(false));
   }, [videoId, get]);

package/dist/_chunks/{index-BAfBs5PQ.js → index-CZeQ6zu2.js}

@@ -130,7 +130,7 @@ const index = {
         defaultMessage: PLUGIN_ID
       },
       Component: async () => {
-        const { App } = await Promise.resolve().then(() => require("./App-Cv1cdLAr.js"));
+        const { App } = await Promise.resolve().then(() => require("./App-57XBOCeK.js"));
         return App;
       }
     });

package/dist/_chunks/{index-K6X5FM2O.mjs → index-CuxGkc_5.mjs}

@@ -129,7 +129,7 @@ const index = {
         defaultMessage: PLUGIN_ID
       },
       Component: async () => {
-        const { App } = await import("./App-bN58O1bN.mjs");
+        const { App } = await import("./App-D7Dy3o44.mjs");
         return App;
       }
     });

package/dist/admin/index.js

@@ -1,4 +1,4 @@
 "use strict";
-const index = require("../_chunks/index-BAfBs5PQ.js");
+const index = require("../_chunks/index-CZeQ6zu2.js");
 require("react/jsx-runtime");
 module.exports = index.index;

package/dist/admin/index.mjs

@@ -1,4 +1,4 @@
-import { i } from "../_chunks/index-K6X5FM2O.mjs";
+import { i } from "../_chunks/index-CuxGkc_5.mjs";
 import "react/jsx-runtime";
 export {
   i as default

package/dist/server/index.js

@@ -1,9 +1,6 @@
 "use strict";
-const openai = require("@langchain/openai");
-const pgvector = require("@langchain/community/vectorstores/pgvector");
-const output_parsers = require("@langchain/core/output_parsers");
-const prompts = require("@langchain/core/prompts");
-const runnables = require("@langchain/core/runnables");
+const ai = require("ai");
+const openai = require("@ai-sdk/openai");
 const pg = require("pg");
 const index_js = require("@modelcontextprotocol/sdk/server/index.js");
 const types_js = require("@modelcontextprotocol/sdk/types.js");
@@ -69,12 +66,11 @@ const config = {
 };
 class PluginManager {
   constructor() {
-    this.embeddings = null;
-    this.chat = null;
+    this.embeddingModel_ = null;
+    this.chatModel = null;
     this.pool = null;
-    this.embeddingModel = "text-embedding-3-small";
+    this.embeddingModelName = "text-embedding-3-small";
     this.dimensions = 1536;
-    this.vectorStoreConfig = null;
   }
   async initializePool(connectionString) {
     console.log("Initializing Neon DB Pool");
@@ -130,74 +126,47 @@ class PluginManager {
       client.release();
     }
   }
-  async initializeEmbeddings(openAIApiKey) {
-    console.log(`Initializing OpenAI Embeddings (model: ${this.embeddingModel})`);
-    if (this.embeddings) return this.embeddings;
-    try {
-      this.embeddings = new openai.OpenAIEmbeddings({
-        openAIApiKey,
-        modelName: this.embeddingModel,
-        dimensions: this.dimensions
-      });
-      return this.embeddings;
-    } catch (error) {
-      console.error(`Failed to initialize Embeddings: ${error}`);
-      throw new Error(`Failed to initialize Embeddings: ${error}`);
-    }
+  initializeEmbeddings(openai2) {
+    console.log(`Initializing OpenAI Embeddings (model: ${this.embeddingModelName})`);
+    if (this.embeddingModel_) return;
+    this.embeddingModel_ = openai2.embedding(this.embeddingModelName, {
+      dimensions: this.dimensions
+    });
   }
-  async initializeChat(openAIApiKey) {
+  initializeChat(openai2) {
     console.log("Initializing Chat Model");
-    if (this.chat) return this.chat;
-    try {
-      this.chat = new openai.ChatOpenAI({
-        modelName: "gpt-4o-mini",
-        temperature: 0.7,
-        openAIApiKey
-      });
-      return this.chat;
-    } catch (error) {
-      console.error(`Failed to initialize Chat: ${error}`);
-      throw new Error(`Failed to initialize Chat: ${error}`);
-    }
+    if (this.chatModel) return;
+    this.chatModel = openai2("gpt-4o-mini");
   }
   async initialize(config2) {
     const model = config2.embeddingModel || "text-embedding-3-small";
     if (EMBEDDING_MODELS[model]) {
-      this.embeddingModel = model;
+      this.embeddingModelName = model;
       this.dimensions = EMBEDDING_MODELS[model].dimensions;
     } else {
       console.warn(`Invalid embedding model "${model}", using default`);
-      this.embeddingModel = "text-embedding-3-small";
+      this.embeddingModelName = "text-embedding-3-small";
       this.dimensions = EMBEDDING_MODELS["text-embedding-3-small"].dimensions;
     }
-    console.log(`Using embedding model: ${this.embeddingModel} (${this.dimensions} dimensions)`);
+    console.log(`Using embedding model: ${this.embeddingModelName} (${this.dimensions} dimensions)`);
     await this.initializePool(config2.neonConnectionString);
-    await this.initializeEmbeddings(config2.openAIApiKey);
-    await this.initializeChat(config2.openAIApiKey);
-    if (this.pool) {
-      this.vectorStoreConfig = {
-        pool: this.pool,
-        tableName: "embeddings_documents",
-        columns: {
-          idColumnName: "id",
-          vectorColumnName: "embedding",
-          contentColumnName: "content",
-          metadataColumnName: "metadata"
-        },
-        distanceStrategy: "cosine"
-      };
-    }
+    const openai$1 = openai.createOpenAI({ apiKey: config2.openAIApiKey });
+    this.initializeEmbeddings(openai$1);
+    this.initializeChat(openai$1);
     console.log("Plugin Manager Initialization Complete");
   }
   async createEmbedding(docData) {
-    if (!this.embeddings || !this.vectorStoreConfig || !this.pool) {
+    if (!this.embeddingModel_ || !this.pool) {
       throw new Error("Plugin manager not initialized");
     }
     const maxRetries = 3;
     const retryDelay = 2e3;
     for (let attempt = 1; attempt <= maxRetries; attempt++) {
       try {
-        const embeddingVector = await this.embeddings.embedQuery(docData.content);
+        const { embedding: embeddingVector } = await ai.embed({
+          model: this.embeddingModel_,
+          value: docData.content
+        });
         const metadata = {
           id: docData.id,
           title: docData.title,
@@ -246,58 +215,56 @@ class PluginManager {
     }
   }
   async queryEmbedding(query) {
-    if (!this.embeddings || !this.chat || !this.vectorStoreConfig) {
+    if (!this.embeddingModel_ || !this.chatModel || !this.pool) {
       throw new Error("Plugin manager not initialized");
     }
     try {
-      const vectorStore = await pgvector.PGVectorStore.initialize(
-        this.embeddings,
-        this.vectorStoreConfig
-      );
-      const resultsWithScores = await vectorStore.similaritySearchWithScore(query, 6);
+      const { embedding: queryVector } = await ai.embed({
+        model: this.embeddingModel_,
+        value: query
+      });
+      const vectorStr = `[${queryVector.join(",")}]`;
+      const results = await this.pool.query(`
+        SELECT
+          content,
+          metadata,
+          1 - (embedding <=> $1::vector) AS similarity
+        FROM embeddings_documents
+        WHERE 1 - (embedding <=> $1::vector) > 0
+        ORDER BY embedding <=> $1::vector
+        LIMIT 6
+      `, [vectorStr]);
       console.log(`[queryEmbedding] Query: "${query}"`);
-      console.log(`[queryEmbedding] Found ${resultsWithScores.length} results:`);
-      resultsWithScores.forEach(([doc, score], i) => {
-        console.log(`  ${i + 1}. Score: ${score.toFixed(4)}, Title: ${doc.metadata?.title || "N/A"}`);
+      console.log(`[queryEmbedding] Found ${results.rows.length} results:`);
+      results.rows.forEach((row, i) => {
+        console.log(`  ${i + 1}. Score: ${row.similarity.toFixed(4)}, Title: ${row.metadata?.title || "N/A"}`);
       });
       const SIMILARITY_THRESHOLD = 1;
-      const relevantResults = resultsWithScores.filter(([_, score]) => score < SIMILARITY_THRESHOLD);
+      const relevantResults = results.rows.filter((row) => row.similarity < SIMILARITY_THRESHOLD);
       console.log(`[queryEmbedding] ${relevantResults.length} results passed threshold (< ${SIMILARITY_THRESHOLD})`);
       const topResults = relevantResults.slice(0, 3);
-      const sourceDocuments = topResults.map(([doc]) => doc);
-      const bestMatchForDisplay = topResults.length > 0 ? [topResults[0][0]] : [];
-      const formatDocs = (docs) => {
-        return docs.map((doc) => {
-          const title = doc.metadata?.title ? `Title: ${doc.metadata.title}
+      const sourceDocuments = topResults.map((row) => ({
+        pageContent: row.content,
+        metadata: row.metadata
+      }));
+      const bestMatchForDisplay = topResults.length > 0 ? [{ pageContent: topResults[0].content, metadata: topResults[0].metadata }] : [];
+      const context = sourceDocuments.map((doc) => {
+        const title = doc.metadata?.title ? `Title: ${doc.metadata.title}
 ` : "";
-          return `${title}${doc.pageContent}`;
-        }).join("\n\n");
-      };
-      const ragPrompt = prompts.ChatPromptTemplate.fromMessages([
-        [
-          "system",
-          `You are a helpful assistant that answers questions based on the provided context.
+        return `${title}${doc.pageContent}`;
+      }).join("\n\n");
+      const { text } = await ai.generateText({
+        model: this.chatModel,
+        system: `You are a helpful assistant that answers questions based on the provided context.
 If you cannot find the answer in the context, say so. Be concise and accurate.
 
 Context:
-{context}`
-        ],
-        ["human", "{question}"]
-      ]);
-      const ragChain = runnables.RunnableSequence.from([
-        {
-          context: async () => formatDocs(sourceDocuments),
-          question: new runnables.RunnablePassthrough()
-        },
-        ragPrompt,
-        this.chat,
-        new output_parsers.StringOutputParser()
-      ]);
-      const text = await ragChain.invoke(query);
+${context}`,
+        prompt: query
+      });
       return {
         text,
         sourceDocuments: bestMatchForDisplay
-        // Only return best match to display
       };
     } catch (error) {
       console.error(`Failed to query embeddings: ${error}`);
@@ -305,22 +272,32 @@ Context:
     }
   }
   async similaritySearch(query, k = 4) {
-    if (!this.embeddings || !this.vectorStoreConfig) {
+    if (!this.embeddingModel_ || !this.pool) {
       throw new Error("Plugin manager not initialized");
     }
     try {
-      const vectorStore = await pgvector.PGVectorStore.initialize(
-        this.embeddings,
-        this.vectorStoreConfig
-      );
-      return await vectorStore.similaritySearch(query, k);
+      const { embedding: queryVector } = await ai.embed({
+        model: this.embeddingModel_,
+        value: query
+      });
+      const vectorStr = `[${queryVector.join(",")}]`;
+      const results = await this.pool.query(`
+        SELECT content, metadata
+        FROM embeddings_documents
+        ORDER BY embedding <=> $1::vector
+        LIMIT $2
+      `, [vectorStr, k]);
+      return results.rows.map((row) => ({
+        pageContent: row.content,
+        metadata: row.metadata
+      }));
     } catch (error) {
       console.error(`Failed to perform similarity search: ${error}`);
       throw new Error(`Failed to perform similarity search: ${error}`);
     }
   }
   isInitialized() {
-    return !!(this.embeddings && this.chat && this.pool);
+    return !!(this.embeddingModel_ && this.chatModel && this.pool);
   }
   /**
    * Get all embeddings from Neon DB
@@ -375,23 +352,19 @@ Context:
   getPool() {
     return this.pool;
   }
-  getEmbeddings() {
-    return this.embeddings;
+  getEmbeddingModel_() {
+    return this.embeddingModel_;
   }
-  getEmbeddingModel() {
-    return this.embeddingModel;
-  }
-  getChat() {
-    return this.chat;
+  getEmbeddingModelName() {
+    return this.embeddingModelName;
   }
   async destroy() {
     if (this.pool) {
       await this.pool.end();
       this.pool = null;
     }
-    this.embeddings = null;
-    this.chat = null;
-    this.vectorStoreConfig = null;
+    this.embeddingModel_ = null;
+    this.chatModel = null;
   }
   /**
    * Clear all embeddings from Neon DB
@@ -454,7 +427,7 @@ const SearchYtKnowledgeSchema = zod.z.object({
   videoId: zod.z.string().optional(),
   topics: zod.z.array(zod.z.string()).optional(),
   contextWindowSeconds: zod.z.number().min(0).optional().default(30),
-  minSimilarity: zod.z.number().min(0).max(1).optional().default(0.65)
+  minSimilarity: zod.z.number().min(0).max(1).optional().default(0.3)
 });
 const GetVideoTranscriptRangeSchema = zod.z.object({
   videoId: zod.z.string().min(1, "Video ID is required"),
@@ -493,7 +466,7 @@ function formatTime$5(seconds) {
 }
 const searchYtKnowledgeMcpTool = {
   name: "search_yt_knowledge",
-  description: "Semantically search YouTube video transcripts. Returns relevant passages with timestamps, deep links, video topics, and summary.",
+  description: "Semantically search YouTube video transcripts. Returns relevant passages with timestamps, deep links, video topics, and summary. IMPORTANT: After receiving results, use the contextText to directly answer the user's question. Cite the video title, timestamp, and deep link. Do not just list results — synthesize an answer from the transcript content.",
   inputSchema: {
     type: "object",
     properties: {
@@ -529,7 +502,7 @@ const searchYtKnowledgeMcpTool = {
 async function handleSearchYtKnowledge(strapi, args) {
   const results = await strapi.plugin("yt-embeddings-strapi-plugin").service("ytEmbeddings").search(args.query, {
     limit: args.limit ?? 5,
-    minSimilarity: args.minSimilarity ?? 0.65,
+    minSimilarity: args.minSimilarity ?? 0.3,
     videoId: args.videoId,
     topics: args.topics,
     contextWindowSeconds: args.contextWindowSeconds ?? 30
@@ -1116,8 +1089,8 @@ Topics: ${(r.topics || []).join(", ")}
 
 ${r.contextText || r.chunkText}`
       ).join("\n\n---\n\n");
-      const chat = pluginManager.getChat();
-      if (!chat) {
+      const config2 = strapi.config.get("plugin::yt-embeddings-strapi-plugin");
+      if (!config2?.openAIApiKey) {
         ctx.body = {
           text: ytResults.map((r) => `**${r.title}** (${r.deepLink})
 ${r.chunkText}`).join("\n\n"),
@@ -1128,19 +1101,19 @@ ${r.chunkText}`).join("\n\n"),
         };
         return;
       }
-      const prompt = prompts.ChatPromptTemplate.fromMessages([
-        ["system", `You are a helpful assistant that answers questions based on YouTube transcript content.
+      const openai$1 = openai.createOpenAI({ apiKey: config2.openAIApiKey });
+      const { text } = await ai.generateText({
+        model: openai$1("gpt-4o-mini"),
+        system: `You are a helpful assistant that answers questions based on YouTube transcript content.
 Include timestamps and video links when relevant. Be concise and accurate.
 If you cannot find the answer in the context, say so.
 
 Context:
-{context}`],
-        ["human", "{question}"]
-      ]);
-      const chain = prompt.pipe(chat);
-      const response = await chain.invoke({ context, question: query });
+${context}`,
+        prompt: query
+      });
       ctx.body = {
-        text: response.content,
+        text,
         sourceDocuments: ytResults.map((r) => ({
           pageContent: r.chunkText,
           metadata: { id: r.videoId, title: r.title, deepLink: r.deepLink }
@@ -1352,7 +1325,7 @@ const ytController = ({ strapi }) => ({
       }
       const result = await strapi.plugin(PLUGIN_ID).service("ytEmbeddings").search(q, {
         limit: limit ? parseInt(limit, 10) : 5,
-        minSimilarity: minSimilarity ? parseFloat(minSimilarity) : 0.65,
+        minSimilarity: minSimilarity ? parseFloat(minSimilarity) : 0.3,
         videoId,
         topics: topics ? topics.split(",") : void 0,
         contextWindowSeconds: contextWindowSeconds ? parseInt(contextWindowSeconds, 10) : 30
@@ -1591,7 +1564,7 @@ function formatTime$2(seconds) {
 }
 const searchYtKnowledgeTool = {
   name: "searchYtKnowledge",
-  description: "Semantically search YouTube video transcripts. Returns relevant passages with timestamps, deep links, video topics, and summary.",
+  description: "Semantically search YouTube video transcripts. Returns relevant passages with timestamps, deep links, video topics, and summary. IMPORTANT: After receiving results, use the contextText to directly answer the user's question. Cite the video title, timestamp, and deep link. Do not just list results — synthesize an answer from the transcript content.",
   schema: SearchYtKnowledgeSchema,
   execute: async (args, strapi) => {
     const validated = SearchYtKnowledgeSchema.parse(args);
@@ -1810,15 +1783,14 @@ const metadataSchema = zod.z.object({
   language: zod.z.string().default("en")
 });
 async function extractVideoMetadata(title, fullTranscript, durationSeconds, openAIApiKey) {
-  const llm = new openai.ChatOpenAI({
-    modelName: "gpt-4o-mini",
-    temperature: 0,
-    openAIApiKey
-  });
-  const structured = llm.withStructuredOutput(metadataSchema);
+  const openai$1 = openai.createOpenAI({ apiKey: openAIApiKey });
   const words = fullTranscript.split(/\s+/);
   const sample = words.length > 4e3 ? [...words.slice(0, 2e3), "...", ...words.slice(-2e3)].join(" ") : fullTranscript;
-  const result = await structured.invoke(`
+  const { object } = await ai.generateObject({
+    model: openai$1("gpt-4o-mini"),
+    schema: metadataSchema,
+    temperature: 0,
+    prompt: `
 Video title: "${title}"
 Duration: ${Math.floor(durationSeconds / 60)} minutes
 
@@ -1832,12 +1804,13 @@ Extract:
 - summary: 2-3 sentences describing what the video teaches or argues
 - keyMoments: the 5-8 most important moments, with approximate start time in seconds
 - language: ISO 639-1 language code of the transcript
-  `.trim());
+    `.trim()
+  });
   return {
-    topics: result.topics ?? [],
-    summary: result.summary ?? "",
-    keyMoments: result.keyMoments ?? [],
-    language: result.language ?? "en"
+    topics: object.topics ?? [],
+    summary: object.summary ?? "",
+    keyMoments: object.keyMoments ?? [],
+    language: object.language ?? "en"
   };
 }
 function computeContentHash(content) {
@@ -1847,9 +1820,9 @@ const ytEmbeddings = ({ strapi }) => ({
   // ── Ingest a single transcript ──────────────────────────────────────────────
   async embedTranscript(transcript) {
     const pool = pluginManager.getPool();
-    const embeddings = pluginManager.getEmbeddings();
-    const embeddingModel = pluginManager.getEmbeddingModel();
-    if (!pool || !embeddings) {
+    const embeddingModel = pluginManager.getEmbeddingModel_();
+    const embeddingModelName = pluginManager.getEmbeddingModelName();
+    if (!pool || !embeddingModel) {
       throw new Error("[yt-embed] Plugin manager not initialized");
     }
     const contentHash = computeContentHash(transcript.fullTranscript);
@@ -1880,7 +1853,7 @@ const ytEmbeddings = ({ strapi }) => ({
         transcript.title,
         durationSeconds,
         contentHash,
-        embeddingModel
+        embeddingModelName
       ]
     );
     try {
@@ -1916,7 +1889,7 @@ const ytEmbeddings = ({ strapi }) => ({
         strapi.log.info(`[yt-embed] ${transcript.title} — no chunks (empty transcript)`);
         return { videoId: transcript.videoId, chunkCount: 0, skipped: false };
       }
-      const embeddingVectors = await embeddings.embedDocuments(chunks.map((c) => c.text));
+      const { embeddings: embeddingVectors } = await ai.embedMany({ model: embeddingModel, values: chunks.map((c) => c.text) });
       const insertedIds = [];
       for (let i = 0; i < chunks.length; i++) {
         const chunk = chunks[i];
@@ -1967,12 +1940,14 @@ const ytEmbeddings = ({ strapi }) => ({
   // ── Semantic search with context expansion ──────────────────────────────────
   async search(query, options = {}) {
     const pool = pluginManager.getPool();
-    const embeddingsClient = pluginManager.getEmbeddings();
-    if (!pool || !embeddingsClient) {
+    const embeddingModel = pluginManager.getEmbeddingModel_();
+    if (!pool || !embeddingModel) {
       throw new Error("[yt-embed] Plugin manager not initialized");
     }
     const { limit = 5, minSimilarity = 0.2, contextWindowSeconds = 30 } = options;
-    const queryVector = await embeddingsClient.embedQuery(query);
+    console.log(`[yt-embed search] Embedding query: "${query}" with model`);
+    const { embedding: queryVector } = await ai.embed({ model: embeddingModel, value: query });
+    console.log(`[yt-embed search] Got embedding vector, length: ${queryVector.length}`);
     const vectorStr = `[${queryVector.join(",")}]`;
     const params = [vectorStr, minSimilarity, limit * 2];
     const filters = [];
@@ -2001,6 +1976,10 @@ const ytEmbeddings = ({ strapi }) => ({
       ORDER BY vc.embedding <=> $1::vector
       LIMIT $3
     `, params);
+    console.log(`[yt-embed search] Query returned ${rows.rows.length} rows (minSimilarity: ${minSimilarity})`);
+    if (rows.rows.length > 0) {
+      console.log(`[yt-embed search] Top similarity: ${rows.rows[0].similarity}`);
+    }
     if (!rows.rows.length) return [];
     const seen = /* @__PURE__ */ new Set();
     const deduped = rows.rows.filter((row) => {

package/dist/server/index.mjs

@@ -1,8 +1,5 @@
-import { OpenAIEmbeddings, ChatOpenAI } from "@langchain/openai";
-import { PGVectorStore } from "@langchain/community/vectorstores/pgvector";
-import { StringOutputParser } from "@langchain/core/output_parsers";
-import { ChatPromptTemplate } from "@langchain/core/prompts";
-import { RunnableSequence, RunnablePassthrough } from "@langchain/core/runnables";
+import { embed, generateText, generateObject, embedMany } from "ai";
+import { createOpenAI } from "@ai-sdk/openai";
 import { Pool } from "pg";
 import { Server } from "@modelcontextprotocol/sdk/server/index.js";
 import { ListToolsRequestSchema, CallToolRequestSchema } from "@modelcontextprotocol/sdk/types.js";
@@ -50,12 +47,11 @@ const config = {
 };
 class PluginManager {
   constructor() {
-    this.embeddings = null;
-    this.chat = null;
+    this.embeddingModel_ = null;
+    this.chatModel = null;
     this.pool = null;
-    this.embeddingModel = "text-embedding-3-small";
+    this.embeddingModelName = "text-embedding-3-small";
     this.dimensions = 1536;
-    this.vectorStoreConfig = null;
   }
   async initializePool(connectionString) {
     console.log("Initializing Neon DB Pool");
@@ -111,74 +107,47 @@ class PluginManager {
       client.release();
     }
   }
-  async initializeEmbeddings(openAIApiKey) {
-    console.log(`Initializing OpenAI Embeddings (model: ${this.embeddingModel})`);
-    if (this.embeddings) return this.embeddings;
-    try {
-      this.embeddings = new OpenAIEmbeddings({
-        openAIApiKey,
-        modelName: this.embeddingModel,
-        dimensions: this.dimensions
-      });
-      return this.embeddings;
-    } catch (error) {
-      console.error(`Failed to initialize Embeddings: ${error}`);
-      throw new Error(`Failed to initialize Embeddings: ${error}`);
-    }
+  initializeEmbeddings(openai) {
+    console.log(`Initializing OpenAI Embeddings (model: ${this.embeddingModelName})`);
+    if (this.embeddingModel_) return;
+    this.embeddingModel_ = openai.embedding(this.embeddingModelName, {
+      dimensions: this.dimensions
+    });
   }
-  async initializeChat(openAIApiKey) {
+  initializeChat(openai) {
     console.log("Initializing Chat Model");
-    if (this.chat) return this.chat;
-    try {
-      this.chat = new ChatOpenAI({
-        modelName: "gpt-4o-mini",
-        temperature: 0.7,
-        openAIApiKey
-      });
-      return this.chat;
-    } catch (error) {
-      console.error(`Failed to initialize Chat: ${error}`);
-      throw new Error(`Failed to initialize Chat: ${error}`);
-    }
+    if (this.chatModel) return;
+    this.chatModel = openai("gpt-4o-mini");
   }
   async initialize(config2) {
     const model = config2.embeddingModel || "text-embedding-3-small";
     if (EMBEDDING_MODELS[model]) {
-      this.embeddingModel = model;
+      this.embeddingModelName = model;
       this.dimensions = EMBEDDING_MODELS[model].dimensions;
     } else {
       console.warn(`Invalid embedding model "${model}", using default`);
-      this.embeddingModel = "text-embedding-3-small";
+      this.embeddingModelName = "text-embedding-3-small";
       this.dimensions = EMBEDDING_MODELS["text-embedding-3-small"].dimensions;
     }
-    console.log(`Using embedding model: ${this.embeddingModel} (${this.dimensions} dimensions)`);
+    console.log(`Using embedding model: ${this.embeddingModelName} (${this.dimensions} dimensions)`);
     await this.initializePool(config2.neonConnectionString);
-    await this.initializeEmbeddings(config2.openAIApiKey);
-    await this.initializeChat(config2.openAIApiKey);
-    if (this.pool) {
-      this.vectorStoreConfig = {
-        pool: this.pool,
-        tableName: "embeddings_documents",
-        columns: {
-          idColumnName: "id",
-          vectorColumnName: "embedding",
-          contentColumnName: "content",
-          metadataColumnName: "metadata"
-        },
-        distanceStrategy: "cosine"
-      };
-    }
+    const openai = createOpenAI({ apiKey: config2.openAIApiKey });
+    this.initializeEmbeddings(openai);
+    this.initializeChat(openai);
     console.log("Plugin Manager Initialization Complete");
   }
   async createEmbedding(docData) {
-    if (!this.embeddings || !this.vectorStoreConfig || !this.pool) {
+    if (!this.embeddingModel_ || !this.pool) {
       throw new Error("Plugin manager not initialized");
     }
     const maxRetries = 3;
     const retryDelay = 2e3;
     for (let attempt = 1; attempt <= maxRetries; attempt++) {
       try {
-        const embeddingVector = await this.embeddings.embedQuery(docData.content);
+        const { embedding: embeddingVector } = await embed({
+          model: this.embeddingModel_,
+          value: docData.content
+        });
         const metadata = {
           id: docData.id,
           title: docData.title,
@@ -227,58 +196,56 @@ class PluginManager {
     }
   }
   async queryEmbedding(query) {
-    if (!this.embeddings || !this.chat || !this.vectorStoreConfig) {
+    if (!this.embeddingModel_ || !this.chatModel || !this.pool) {
       throw new Error("Plugin manager not initialized");
     }
     try {
-      const vectorStore = await PGVectorStore.initialize(
-        this.embeddings,
-        this.vectorStoreConfig
-      );
-      const resultsWithScores = await vectorStore.similaritySearchWithScore(query, 6);
+      const { embedding: queryVector } = await embed({
+        model: this.embeddingModel_,
+        value: query
+      });
+      const vectorStr = `[${queryVector.join(",")}]`;
+      const results = await this.pool.query(`
+        SELECT
+          content,
+          metadata,
+          1 - (embedding <=> $1::vector) AS similarity
+        FROM embeddings_documents
+        WHERE 1 - (embedding <=> $1::vector) > 0
+        ORDER BY embedding <=> $1::vector
+        LIMIT 6
+      `, [vectorStr]);
       console.log(`[queryEmbedding] Query: "${query}"`);
-      console.log(`[queryEmbedding] Found ${resultsWithScores.length} results:`);
-      resultsWithScores.forEach(([doc, score], i) => {
-        console.log(`  ${i + 1}. Score: ${score.toFixed(4)}, Title: ${doc.metadata?.title || "N/A"}`);
+      console.log(`[queryEmbedding] Found ${results.rows.length} results:`);
+      results.rows.forEach((row, i) => {
+        console.log(`  ${i + 1}. Score: ${row.similarity.toFixed(4)}, Title: ${row.metadata?.title || "N/A"}`);
       });
       const SIMILARITY_THRESHOLD = 1;
-      const relevantResults = resultsWithScores.filter(([_, score]) => score < SIMILARITY_THRESHOLD);
+      const relevantResults = results.rows.filter((row) => row.similarity < SIMILARITY_THRESHOLD);
       console.log(`[queryEmbedding] ${relevantResults.length} results passed threshold (< ${SIMILARITY_THRESHOLD})`);
       const topResults = relevantResults.slice(0, 3);
-      const sourceDocuments = topResults.map(([doc]) => doc);
-      const bestMatchForDisplay = topResults.length > 0 ? [topResults[0][0]] : [];
-      const formatDocs = (docs) => {
-        return docs.map((doc) => {
-          const title = doc.metadata?.title ? `Title: ${doc.metadata.title}
+      const sourceDocuments = topResults.map((row) => ({
+        pageContent: row.content,
+        metadata: row.metadata
+      }));
+      const bestMatchForDisplay = topResults.length > 0 ? [{ pageContent: topResults[0].content, metadata: topResults[0].metadata }] : [];
+      const context = sourceDocuments.map((doc) => {
+        const title = doc.metadata?.title ? `Title: ${doc.metadata.title}
 ` : "";
-          return `${title}${doc.pageContent}`;
-        }).join("\n\n");
-      };
-      const ragPrompt = ChatPromptTemplate.fromMessages([
-        [
-          "system",
-          `You are a helpful assistant that answers questions based on the provided context.
+        return `${title}${doc.pageContent}`;
+      }).join("\n\n");
+      const { text } = await generateText({
+        model: this.chatModel,
+        system: `You are a helpful assistant that answers questions based on the provided context.
 If you cannot find the answer in the context, say so. Be concise and accurate.
 
 Context:
-{context}`
-        ],
-        ["human", "{question}"]
-      ]);
-      const ragChain = RunnableSequence.from([
-        {
-          context: async () => formatDocs(sourceDocuments),
-          question: new RunnablePassthrough()
-        },
-        ragPrompt,
-        this.chat,
-        new StringOutputParser()
-      ]);
-      const text = await ragChain.invoke(query);
+${context}`,
+        prompt: query
+      });
       return {
         text,
         sourceDocuments: bestMatchForDisplay
-        // Only return best match to display
       };
     } catch (error) {
       console.error(`Failed to query embeddings: ${error}`);
@@ -286,22 +253,32 @@ Context:
     }
   }
   async similaritySearch(query, k = 4) {
-    if (!this.embeddings || !this.vectorStoreConfig) {
+    if (!this.embeddingModel_ || !this.pool) {
       throw new Error("Plugin manager not initialized");
     }
     try {
-      const vectorStore = await PGVectorStore.initialize(
-        this.embeddings,
-        this.vectorStoreConfig
-      );
-      return await vectorStore.similaritySearch(query, k);
+      const { embedding: queryVector } = await embed({
+        model: this.embeddingModel_,
+        value: query
+      });
+      const vectorStr = `[${queryVector.join(",")}]`;
+      const results = await this.pool.query(`
+        SELECT content, metadata
+        FROM embeddings_documents
+        ORDER BY embedding <=> $1::vector
+        LIMIT $2
+      `, [vectorStr, k]);
+      return results.rows.map((row) => ({
+        pageContent: row.content,
+        metadata: row.metadata
+      }));
     } catch (error) {
       console.error(`Failed to perform similarity search: ${error}`);
       throw new Error(`Failed to perform similarity search: ${error}`);
     }
   }
   isInitialized() {
-    return !!(this.embeddings && this.chat && this.pool);
+    return !!(this.embeddingModel_ && this.chatModel && this.pool);
   }
   /**
    * Get all embeddings from Neon DB
@@ -356,23 +333,19 @@ Context:
   getPool() {
     return this.pool;
   }
-  getEmbeddings() {
-    return this.embeddings;
+  getEmbeddingModel_() {
+    return this.embeddingModel_;
   }
-  getEmbeddingModel() {
-    return this.embeddingModel;
-  }
-  getChat() {
-    return this.chat;
+  getEmbeddingModelName() {
+    return this.embeddingModelName;
   }
   async destroy() {
     if (this.pool) {
       await this.pool.end();
       this.pool = null;
     }
-    this.embeddings = null;
-    this.chat = null;
-    this.vectorStoreConfig = null;
+    this.embeddingModel_ = null;
+    this.chatModel = null;
   }
   /**
    * Clear all embeddings from Neon DB
@@ -435,7 +408,7 @@ const SearchYtKnowledgeSchema = z.object({
   videoId: z.string().optional(),
   topics: z.array(z.string()).optional(),
   contextWindowSeconds: z.number().min(0).optional().default(30),
-  minSimilarity: z.number().min(0).max(1).optional().default(0.65)
+  minSimilarity: z.number().min(0).max(1).optional().default(0.3)
 });
 const GetVideoTranscriptRangeSchema = z.object({
   videoId: z.string().min(1, "Video ID is required"),
@@ -474,7 +447,7 @@ function formatTime$5(seconds) {
 }
 const searchYtKnowledgeMcpTool = {
   name: "search_yt_knowledge",
-  description: "Semantically search YouTube video transcripts. Returns relevant passages with timestamps, deep links, video topics, and summary.",
+  description: "Semantically search YouTube video transcripts. Returns relevant passages with timestamps, deep links, video topics, and summary. IMPORTANT: After receiving results, use the contextText to directly answer the user's question. Cite the video title, timestamp, and deep link. Do not just list results — synthesize an answer from the transcript content.",
   inputSchema: {
     type: "object",
     properties: {
@@ -510,7 +483,7 @@ const searchYtKnowledgeMcpTool = {
 async function handleSearchYtKnowledge(strapi, args) {
   const results = await strapi.plugin("yt-embeddings-strapi-plugin").service("ytEmbeddings").search(args.query, {
     limit: args.limit ?? 5,
-    minSimilarity: args.minSimilarity ?? 0.65,
+    minSimilarity: args.minSimilarity ?? 0.3,
     videoId: args.videoId,
     topics: args.topics,
     contextWindowSeconds: args.contextWindowSeconds ?? 30
@@ -1097,8 +1070,8 @@ Topics: ${(r.topics || []).join(", ")}
 
 ${r.contextText || r.chunkText}`
       ).join("\n\n---\n\n");
-      const chat = pluginManager.getChat();
-      if (!chat) {
+      const config2 = strapi.config.get("plugin::yt-embeddings-strapi-plugin");
+      if (!config2?.openAIApiKey) {
         ctx.body = {
           text: ytResults.map((r) => `**${r.title}** (${r.deepLink})
 ${r.chunkText}`).join("\n\n"),
@@ -1109,19 +1082,19 @@ ${r.chunkText}`).join("\n\n"),
         };
         return;
       }
-      const prompt = ChatPromptTemplate.fromMessages([
-        ["system", `You are a helpful assistant that answers questions based on YouTube transcript content.
+      const openai = createOpenAI({ apiKey: config2.openAIApiKey });
+      const { text } = await generateText({
+        model: openai("gpt-4o-mini"),
+        system: `You are a helpful assistant that answers questions based on YouTube transcript content.
 Include timestamps and video links when relevant. Be concise and accurate.
 If you cannot find the answer in the context, say so.
 
 Context:
-{context}`],
-        ["human", "{question}"]
-      ]);
-      const chain = prompt.pipe(chat);
-      const response = await chain.invoke({ context, question: query });
+${context}`,
+        prompt: query
+      });
       ctx.body = {
-        text: response.content,
+        text,
         sourceDocuments: ytResults.map((r) => ({
           pageContent: r.chunkText,
           metadata: { id: r.videoId, title: r.title, deepLink: r.deepLink }
@@ -1333,7 +1306,7 @@ const ytController = ({ strapi }) => ({
       }
       const result = await strapi.plugin(PLUGIN_ID).service("ytEmbeddings").search(q, {
         limit: limit ? parseInt(limit, 10) : 5,
-        minSimilarity: minSimilarity ? parseFloat(minSimilarity) : 0.65,
+        minSimilarity: minSimilarity ? parseFloat(minSimilarity) : 0.3,
         videoId,
         topics: topics ? topics.split(",") : void 0,
         contextWindowSeconds: contextWindowSeconds ? parseInt(contextWindowSeconds, 10) : 30
@@ -1572,7 +1545,7 @@ function formatTime$2(seconds) {
 }
 const searchYtKnowledgeTool = {
   name: "searchYtKnowledge",
-  description: "Semantically search YouTube video transcripts. Returns relevant passages with timestamps, deep links, video topics, and summary.",
+  description: "Semantically search YouTube video transcripts. Returns relevant passages with timestamps, deep links, video topics, and summary. IMPORTANT: After receiving results, use the contextText to directly answer the user's question. Cite the video title, timestamp, and deep link. Do not just list results — synthesize an answer from the transcript content.",
   schema: SearchYtKnowledgeSchema,
   execute: async (args, strapi) => {
     const validated = SearchYtKnowledgeSchema.parse(args);
@@ -1791,15 +1764,14 @@ const metadataSchema = z.object({
   language: z.string().default("en")
 });
 async function extractVideoMetadata(title, fullTranscript, durationSeconds, openAIApiKey) {
-  const llm = new ChatOpenAI({
-    modelName: "gpt-4o-mini",
-    temperature: 0,
-    openAIApiKey
-  });
-  const structured = llm.withStructuredOutput(metadataSchema);
+  const openai = createOpenAI({ apiKey: openAIApiKey });
   const words = fullTranscript.split(/\s+/);
   const sample = words.length > 4e3 ? [...words.slice(0, 2e3), "...", ...words.slice(-2e3)].join(" ") : fullTranscript;
-  const result = await structured.invoke(`
+  const { object } = await generateObject({
+    model: openai("gpt-4o-mini"),
+    schema: metadataSchema,
+    temperature: 0,
+    prompt: `
 Video title: "${title}"
 Duration: ${Math.floor(durationSeconds / 60)} minutes
 
@@ -1813,12 +1785,13 @@ Extract:
 - summary: 2-3 sentences describing what the video teaches or argues
 - keyMoments: the 5-8 most important moments, with approximate start time in seconds
 - language: ISO 639-1 language code of the transcript
-  `.trim());
+    `.trim()
+  });
   return {
-    topics: result.topics ?? [],
-    summary: result.summary ?? "",
-    keyMoments: result.keyMoments ?? [],
-    language: result.language ?? "en"
+    topics: object.topics ?? [],
+    summary: object.summary ?? "",
+    keyMoments: object.keyMoments ?? [],
+    language: object.language ?? "en"
   };
 }
 function computeContentHash(content) {
@@ -1828,9 +1801,9 @@ const ytEmbeddings = ({ strapi }) => ({
   // ── Ingest a single transcript ──────────────────────────────────────────────
   async embedTranscript(transcript) {
     const pool = pluginManager.getPool();
-    const embeddings = pluginManager.getEmbeddings();
-    const embeddingModel = pluginManager.getEmbeddingModel();
-    if (!pool || !embeddings) {
+    const embeddingModel = pluginManager.getEmbeddingModel_();
+    const embeddingModelName = pluginManager.getEmbeddingModelName();
+    if (!pool || !embeddingModel) {
       throw new Error("[yt-embed] Plugin manager not initialized");
     }
     const contentHash = computeContentHash(transcript.fullTranscript);
@@ -1861,7 +1834,7 @@ const ytEmbeddings = ({ strapi }) => ({
         transcript.title,
         durationSeconds,
         contentHash,
-        embeddingModel
+        embeddingModelName
       ]
     );
     try {
@@ -1897,7 +1870,7 @@ const ytEmbeddings = ({ strapi }) => ({
         strapi.log.info(`[yt-embed] ${transcript.title} — no chunks (empty transcript)`);
         return { videoId: transcript.videoId, chunkCount: 0, skipped: false };
       }
-      const embeddingVectors = await embeddings.embedDocuments(chunks.map((c) => c.text));
+      const { embeddings: embeddingVectors } = await embedMany({ model: embeddingModel, values: chunks.map((c) => c.text) });
       const insertedIds = [];
       for (let i = 0; i < chunks.length; i++) {
         const chunk = chunks[i];
@@ -1948,12 +1921,14 @@ const ytEmbeddings = ({ strapi }) => ({
   // ── Semantic search with context expansion ──────────────────────────────────
   async search(query, options = {}) {
     const pool = pluginManager.getPool();
-    const embeddingsClient = pluginManager.getEmbeddings();
-    if (!pool || !embeddingsClient) {
+    const embeddingModel = pluginManager.getEmbeddingModel_();
+    if (!pool || !embeddingModel) {
       throw new Error("[yt-embed] Plugin manager not initialized");
     }
     const { limit = 5, minSimilarity = 0.2, contextWindowSeconds = 30 } = options;
-    const queryVector = await embeddingsClient.embedQuery(query);
+    console.log(`[yt-embed search] Embedding query: "${query}" with model`);
+    const { embedding: queryVector } = await embed({ model: embeddingModel, value: query });
+    console.log(`[yt-embed search] Got embedding vector, length: ${queryVector.length}`);
     const vectorStr = `[${queryVector.join(",")}]`;
     const params = [vectorStr, minSimilarity, limit * 2];
     const filters = [];
@@ -1982,6 +1957,10 @@ const ytEmbeddings = ({ strapi }) => ({
       ORDER BY vc.embedding <=> $1::vector
       LIMIT $3
     `, params);
+    console.log(`[yt-embed search] Query returned ${rows.rows.length} rows (minSimilarity: ${minSimilarity})`);
+    if (rows.rows.length > 0) {
+      console.log(`[yt-embed search] Top similarity: ${rows.rows[0].similarity}`);
+    }
     if (!rows.rows.length) return [];
     const seen = /* @__PURE__ */ new Set();
     const deduped = rows.rows.filter((row) => {

package/dist/server/src/plugin-manager.d.ts

@@ -1,5 +1,4 @@
-import { OpenAIEmbeddings, ChatOpenAI } from "@langchain/openai";
-import { Document } from "@langchain/core/documents";
+import type { EmbeddingModel } from "ai";
 import { Pool } from "pg";
 import { type EmbeddingModelName } from "./config";
 interface PluginConfig {
@@ -20,24 +19,29 @@ interface CreateEmbeddingResult {
 }
 interface QueryResponse {
     text: string;
-    sourceDocuments: Document[];
+    sourceDocuments: Array<{
+        pageContent: string;
+        metadata: any;
+    }>;
 }
 declare class PluginManager {
-    private embeddings;
-    private chat;
+    private embeddingModel_;
+    private chatModel;
     private pool;
-    private embeddingModel;
+    private embeddingModelName;
     private dimensions;
-    private vectorStoreConfig;
     initializePool(connectionString: string): Promise<Pool>;
     private initializeVectorTable;
-    initializeEmbeddings(openAIApiKey: string): Promise<OpenAIEmbeddings>;
-    initializeChat(openAIApiKey: string): Promise<ChatOpenAI>;
+    private initializeEmbeddings;
+    private initializeChat;
     initialize(config: PluginConfig): Promise<void>;
     createEmbedding(docData: EmbeddingDocument): Promise<CreateEmbeddingResult>;
     deleteEmbedding(strapiId: string): Promise<void>;
     queryEmbedding(query: string): Promise<QueryResponse>;
-    similaritySearch(query: string, k?: number): Promise<Document[]>;
+    similaritySearch(query: string, k?: number): Promise<Array<{
+        pageContent: string;
+        metadata: any;
+    }>>;
     isInitialized(): boolean;
     /**
      * Get all embeddings from Neon DB
@@ -56,9 +60,8 @@ declare class PluginManager {
      */
     deleteNeonEmbeddingById(neonId: string): Promise<void>;
     getPool(): Pool | null;
-    getEmbeddings(): OpenAIEmbeddings | null;
-    getEmbeddingModel(): EmbeddingModelName;
-    getChat(): ChatOpenAI | null;
+    getEmbeddingModel_(): EmbeddingModel<string> | null;
+    getEmbeddingModelName(): EmbeddingModelName;
     destroy(): Promise<void>;
     /**
      * Clear all embeddings from Neon DB

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "yt-embeddings-strapi-plugin",
-  "version": "0.0.1",
+  "version": "0.1.1",
   "description": "Strapi v5 plugin for vector embeddings with OpenAI and Neon PostgreSQL. Enables semantic search, RAG chat, and MCP (Model Context Protocol) integration.",
   "keywords": [
     "strapi",
@@ -25,9 +25,6 @@
     "url": "https://github.com/PaulBratslavsky/strapi-content-embeddings/issues"
   },
   "type": "commonjs",
-  "overrides": {
-    "@langchain/core": "^1.1.31"
-  },
   "exports": {
     "./package.json": "./package.json",
     "./strapi-admin": {
@@ -58,15 +55,13 @@
     "watch:link": "strapi-plugin watch:link"
   },
   "dependencies": {
-    "@langchain/community": "^1.1.2",
-    "@langchain/core": "^1.1.31",
-    "@langchain/openai": "^1.2.1",
+    "@ai-sdk/openai": "^1.x",
     "@mdxeditor/editor": "^3.52.3",
     "@modelcontextprotocol/sdk": "^1.12.0",
     "@strapi/design-system": "^2.0.0-rc.12",
     "@strapi/icons": "^2.0.0-rc.12",
+    "ai": "^4.x",
     "html-to-text": "^9.0.5",
-    "langchain": "^1.2.4",
     "pg": "^8.13.1",
     "remove-markdown": "^0.5.5",
     "qs": "^6.13.1",