Do not go through LLM to embed when embedding documents (#1428)

server/utils/helpers/index.js

@@ -100,6 +100,7 @@ function getLLMProvider({ provider = null, model = null } = {}) {
 }
 
 function getEmbeddingEngineSelection() {
+  const { NativeEmbedder } = require("../EmbeddingEngines/native");
   const engineSelection = process.env.EMBEDDING_ENGINE;
   switch (engineSelection) {
     case "openai":
@@ -117,7 +118,6 @@ function getEmbeddingEngineSelection() {
       const { OllamaEmbedder } = require("../EmbeddingEngines/ollama");
       return new OllamaEmbedder();
     case "native":
-      const { NativeEmbedder } = require("../EmbeddingEngines/native");
       return new NativeEmbedder();
     case "lmstudio":
       const { LMStudioEmbedder } = require("../EmbeddingEngines/lmstudio");
@@ -126,7 +126,7 @@ function getEmbeddingEngineSelection() {
       const { CohereEmbedder } = require("../EmbeddingEngines/cohere");
       return new CohereEmbedder();
     default:
-      return null;
+      return new NativeEmbedder();
   }
 }
 

server/utils/vectorDbProviders/astra/index.js

@@ -3,11 +3,7 @@ const { TextSplitter } = require("../../TextSplitter");
 const { SystemSettings } = require("../../../models/systemSettings");
 const { storeVectorResult, cachedVectorInformation } = require("../../files");
 const { v4: uuidv4 } = require("uuid");
-const {
-  toChunks,
-  getLLMProvider,
-  getEmbeddingEngineSelection,
-} = require("../../helpers");
+const { toChunks, getEmbeddingEngineSelection } = require("../../helpers");
 const { sourceIdentifier } = require("../../chats");
 
 const AstraDB = {
@@ -149,12 +145,13 @@ const AstraDB = {
         return { vectorized: true, error: null };
       }
 
+      const EmbedderEngine = getEmbeddingEngineSelection();
       const textSplitter = new TextSplitter({
         chunkSize: TextSplitter.determineMaxChunkSize(
           await SystemSettings.getValueOrFallback({
             label: "text_splitter_chunk_size",
           }),
-          getEmbeddingEngineSelection()?.embeddingMaxChunkLength
+          EmbedderEngine?.embeddingMaxChunkLength
         ),
         chunkOverlap: await SystemSettings.getValueOrFallback(
           { label: "text_splitter_chunk_overlap" },
@@ -164,10 +161,9 @@ const AstraDB = {
       const textChunks = await textSplitter.splitText(pageContent);
 
       console.log("Chunks created from document:", textChunks.length);
-      const LLMConnector = getLLMProvider();
       const documentVectors = [];
       const vectors = [];
-      const vectorValues = await LLMConnector.embedChunks(textChunks);
+      const vectorValues = await EmbedderEngine.embedChunks(textChunks);
 
       if (!!vectorValues && vectorValues.length > 0) {
         for (const [i, vector] of vectorValues.entries()) {

server/utils/vectorDbProviders/chroma/index.js

@@ -3,11 +3,7 @@ const { TextSplitter } = require("../../TextSplitter");
 const { SystemSettings } = require("../../../models/systemSettings");
 const { storeVectorResult, cachedVectorInformation } = require("../../files");
 const { v4: uuidv4 } = require("uuid");
-const {
-  toChunks,
-  getLLMProvider,
-  getEmbeddingEngineSelection,
-} = require("../../helpers");
+const { toChunks, getEmbeddingEngineSelection } = require("../../helpers");
 const { parseAuthHeader } = require("../../http");
 const { sourceIdentifier } = require("../../chats");
 
@@ -192,12 +188,13 @@ const Chroma = {
       // We have to do this manually as opposed to using LangChains `Chroma.fromDocuments`
       // because we then cannot atomically control our namespace to granularly find/remove documents
       // from vectordb.
+      const EmbedderEngine = getEmbeddingEngineSelection();
       const textSplitter = new TextSplitter({
         chunkSize: TextSplitter.determineMaxChunkSize(
           await SystemSettings.getValueOrFallback({
             label: "text_splitter_chunk_size",
           }),
-          getEmbeddingEngineSelection()?.embeddingMaxChunkLength
+          EmbedderEngine?.embeddingMaxChunkLength
         ),
         chunkOverlap: await SystemSettings.getValueOrFallback(
           { label: "text_splitter_chunk_overlap" },
@@ -207,10 +204,9 @@ const Chroma = {
       const textChunks = await textSplitter.splitText(pageContent);
 
       console.log("Chunks created from document:", textChunks.length);
-      const LLMConnector = getLLMProvider();
       const documentVectors = [];
       const vectors = [];
-      const vectorValues = await LLMConnector.embedChunks(textChunks);
+      const vectorValues = await EmbedderEngine.embedChunks(textChunks);
       const submission = {
         ids: [],
         embeddings: [],

server/utils/vectorDbProviders/lance/index.js

@@ -1,9 +1,5 @@
 const lancedb = require("vectordb");
-const {
-  toChunks,
-  getLLMProvider,
-  getEmbeddingEngineSelection,
-} = require("../../helpers");
+const { toChunks, getEmbeddingEngineSelection } = require("../../helpers");
 const { TextSplitter } = require("../../TextSplitter");
 const { SystemSettings } = require("../../../models/systemSettings");
 const { storeVectorResult, cachedVectorInformation } = require("../../files");
@@ -190,12 +186,13 @@ const LanceDb = {
       // We have to do this manually as opposed to using LangChains `xyz.fromDocuments`
       // because we then cannot atomically control our namespace to granularly find/remove documents
       // from vectordb.
+      const EmbedderEngine = getEmbeddingEngineSelection();
       const textSplitter = new TextSplitter({
         chunkSize: TextSplitter.determineMaxChunkSize(
           await SystemSettings.getValueOrFallback({
             label: "text_splitter_chunk_size",
           }),
-          getEmbeddingEngineSelection()?.embeddingMaxChunkLength
+          EmbedderEngine?.embeddingMaxChunkLength
         ),
         chunkOverlap: await SystemSettings.getValueOrFallback(
           { label: "text_splitter_chunk_overlap" },
@@ -205,11 +202,10 @@ const LanceDb = {
       const textChunks = await textSplitter.splitText(pageContent);
 
       console.log("Chunks created from document:", textChunks.length);
-      const LLMConnector = getLLMProvider();
       const documentVectors = [];
       const vectors = [];
       const submissions = [];
-      const vectorValues = await LLMConnector.embedChunks(textChunks);
+      const vectorValues = await EmbedderEngine.embedChunks(textChunks);
 
       if (!!vectorValues && vectorValues.length > 0) {
         for (const [i, vector] of vectorValues.entries()) {

server/utils/vectorDbProviders/milvus/index.js

@@ -8,11 +8,7 @@ const { TextSplitter } = require("../../TextSplitter");
 const { SystemSettings } = require("../../../models/systemSettings");
 const { v4: uuidv4 } = require("uuid");
 const { storeVectorResult, cachedVectorInformation } = require("../../files");
-const {
-  toChunks,
-  getLLMProvider,
-  getEmbeddingEngineSelection,
-} = require("../../helpers");
+const { toChunks, getEmbeddingEngineSelection } = require("../../helpers");
 const { sourceIdentifier } = require("../../chats");
 
 const Milvus = {
@@ -184,12 +180,13 @@ const Milvus = {
         return { vectorized: true, error: null };
       }
 
+      const EmbedderEngine = getEmbeddingEngineSelection();
       const textSplitter = new TextSplitter({
         chunkSize: TextSplitter.determineMaxChunkSize(
           await SystemSettings.getValueOrFallback({
             label: "text_splitter_chunk_size",
           }),
-          getEmbeddingEngineSelection()?.embeddingMaxChunkLength
+          EmbedderEngine?.embeddingMaxChunkLength
         ),
         chunkOverlap: await SystemSettings.getValueOrFallback(
           { label: "text_splitter_chunk_overlap" },
@@ -199,10 +196,9 @@ const Milvus = {
       const textChunks = await textSplitter.splitText(pageContent);
 
       console.log("Chunks created from document:", textChunks.length);
-      const LLMConnector = getLLMProvider();
       const documentVectors = [];
       const vectors = [];
-      const vectorValues = await LLMConnector.embedChunks(textChunks);
+      const vectorValues = await EmbedderEngine.embedChunks(textChunks);
 
       if (!!vectorValues && vectorValues.length > 0) {
         for (const [i, vector] of vectorValues.entries()) {

server/utils/vectorDbProviders/pinecone/index.js

@@ -3,11 +3,7 @@ const { TextSplitter } = require("../../TextSplitter");
 const { SystemSettings } = require("../../../models/systemSettings");
 const { storeVectorResult, cachedVectorInformation } = require("../../files");
 const { v4: uuidv4 } = require("uuid");
-const {
-  toChunks,
-  getLLMProvider,
-  getEmbeddingEngineSelection,
-} = require("../../helpers");
+const { toChunks, getEmbeddingEngineSelection } = require("../../helpers");
 const { sourceIdentifier } = require("../../chats");
 
 const PineconeDB = {
@@ -135,12 +131,13 @@ const PineconeDB = {
       // because we then cannot atomically control our namespace to granularly find/remove documents
       // from vectordb.
       // https://github.com/hwchase17/langchainjs/blob/2def486af734c0ca87285a48f1a04c057ab74bdf/langchain/src/vectorstores/pinecone.ts#L167
+      const EmbedderEngine = getEmbeddingEngineSelection();
       const textSplitter = new TextSplitter({
         chunkSize: TextSplitter.determineMaxChunkSize(
           await SystemSettings.getValueOrFallback({
             label: "text_splitter_chunk_size",
           }),
-          getEmbeddingEngineSelection()?.embeddingMaxChunkLength
+          EmbedderEngine?.embeddingMaxChunkLength
         ),
         chunkOverlap: await SystemSettings.getValueOrFallback(
           { label: "text_splitter_chunk_overlap" },
@@ -150,10 +147,9 @@ const PineconeDB = {
       const textChunks = await textSplitter.splitText(pageContent);
 
       console.log("Chunks created from document:", textChunks.length);
-      const LLMConnector = getLLMProvider();
       const documentVectors = [];
       const vectors = [];
-      const vectorValues = await LLMConnector.embedChunks(textChunks);
+      const vectorValues = await EmbedderEngine.embedChunks(textChunks);
 
       if (!!vectorValues && vectorValues.length > 0) {
         for (const [i, vector] of vectorValues.entries()) {

server/utils/vectorDbProviders/qdrant/index.js

@@ -3,11 +3,7 @@ const { TextSplitter } = require("../../TextSplitter");
 const { SystemSettings } = require("../../../models/systemSettings");
 const { storeVectorResult, cachedVectorInformation } = require("../../files");
 const { v4: uuidv4 } = require("uuid");
-const {
-  toChunks,
-  getLLMProvider,
-  getEmbeddingEngineSelection,
-} = require("../../helpers");
+const { toChunks, getEmbeddingEngineSelection } = require("../../helpers");
 const { sourceIdentifier } = require("../../chats");
 
 const QDrant = {
@@ -209,12 +205,13 @@ const QDrant = {
       // We have to do this manually as opposed to using LangChains `Qdrant.fromDocuments`
       // because we then cannot atomically control our namespace to granularly find/remove documents
       // from vectordb.
+      const EmbedderEngine = getEmbeddingEngineSelection();
       const textSplitter = new TextSplitter({
         chunkSize: TextSplitter.determineMaxChunkSize(
           await SystemSettings.getValueOrFallback({
             label: "text_splitter_chunk_size",
           }),
-          getEmbeddingEngineSelection()?.embeddingMaxChunkLength
+          EmbedderEngine?.embeddingMaxChunkLength
         ),
         chunkOverlap: await SystemSettings.getValueOrFallback(
           { label: "text_splitter_chunk_overlap" },
@@ -224,10 +221,9 @@ const QDrant = {
       const textChunks = await textSplitter.splitText(pageContent);
 
       console.log("Chunks created from document:", textChunks.length);
-      const LLMConnector = getLLMProvider();
       const documentVectors = [];
       const vectors = [];
-      const vectorValues = await LLMConnector.embedChunks(textChunks);
+      const vectorValues = await EmbedderEngine.embedChunks(textChunks);
       const submission = {
         ids: [],
         vectors: [],

server/utils/vectorDbProviders/weaviate/index.js

@@ -3,11 +3,7 @@ const { TextSplitter } = require("../../TextSplitter");
 const { SystemSettings } = require("../../../models/systemSettings");
 const { storeVectorResult, cachedVectorInformation } = require("../../files");
 const { v4: uuidv4 } = require("uuid");
-const {
-  toChunks,
-  getLLMProvider,
-  getEmbeddingEngineSelection,
-} = require("../../helpers");
+const { toChunks, getEmbeddingEngineSelection } = require("../../helpers");
 const { camelCase } = require("../../helpers/camelcase");
 const { sourceIdentifier } = require("../../chats");
 
@@ -251,12 +247,13 @@ const Weaviate = {
       // We have to do this manually as opposed to using LangChains `Chroma.fromDocuments`
       // because we then cannot atomically control our namespace to granularly find/remove documents
       // from vectordb.
+      const EmbedderEngine = getEmbeddingEngineSelection();
       const textSplitter = new TextSplitter({
         chunkSize: TextSplitter.determineMaxChunkSize(
           await SystemSettings.getValueOrFallback({
             label: "text_splitter_chunk_size",
           }),
-          getEmbeddingEngineSelection()?.embeddingMaxChunkLength
+          EmbedderEngine?.embeddingMaxChunkLength
         ),
         chunkOverlap: await SystemSettings.getValueOrFallback(
           { label: "text_splitter_chunk_overlap" },
@@ -266,10 +263,9 @@ const Weaviate = {
       const textChunks = await textSplitter.splitText(pageContent);
 
       console.log("Chunks created from document:", textChunks.length);
-      const LLMConnector = getLLMProvider();
       const documentVectors = [];
       const vectors = [];
-      const vectorValues = await LLMConnector.embedChunks(textChunks);
+      const vectorValues = await EmbedderEngine.embedChunks(textChunks);
       const submission = {
         ids: [],
         vectors: [],

server/utils/vectorDbProviders/zilliz/index.js

@@ -8,11 +8,7 @@ const { TextSplitter } = require("../../TextSplitter");
 const { SystemSettings } = require("../../../models/systemSettings");
 const { v4: uuidv4 } = require("uuid");
 const { storeVectorResult, cachedVectorInformation } = require("../../files");
-const {
-  toChunks,
-  getLLMProvider,
-  getEmbeddingEngineSelection,
-} = require("../../helpers");
+const { toChunks, getEmbeddingEngineSelection } = require("../../helpers");
 const { sourceIdentifier } = require("../../chats");
 
 // Zilliz is basically a copy of Milvus DB class with a different constructor
@@ -185,12 +181,13 @@ const Zilliz = {
         return { vectorized: true, error: null };
       }
 
+      const EmbedderEngine = getEmbeddingEngineSelection();
       const textSplitter = new TextSplitter({
         chunkSize: TextSplitter.determineMaxChunkSize(
           await SystemSettings.getValueOrFallback({
             label: "text_splitter_chunk_size",
           }),
-          getEmbeddingEngineSelection()?.embeddingMaxChunkLength
+          EmbedderEngine?.embeddingMaxChunkLength
         ),
         chunkOverlap: await SystemSettings.getValueOrFallback(
           { label: "text_splitter_chunk_overlap" },
@@ -200,10 +197,9 @@ const Zilliz = {
       const textChunks = await textSplitter.splitText(pageContent);
 
       console.log("Chunks created from document:", textChunks.length);
-      const LLMConnector = getLLMProvider();
       const documentVectors = [];
       const vectors = [];
-      const vectorValues = await LLMConnector.embedChunks(textChunks);
+      const vectorValues = await EmbedderEngine.embedChunks(textChunks);
 
       if (!!vectorValues && vectorValues.length > 0) {
         for (const [i, vector] of vectorValues.entries()) {