fix(ai-bot): correct configuration to call ai.lies.exposed from remote

Author: ascariandrea

helm/templates/services/localai/localai-env.configmap.yaml

@@ -1,8 +1,10 @@
 apiVersion: v1
 data:
+  DEBUG: "true"
   LOCALAI_LOG_LEVEL: debug
   LOCALAI_THREADS: "4"
   LOCALAI_API_KEY: {{ .Values.localai.openAIApiKey | quote }}
+  LOCALAI_SINGLE_ACTIVE_BACKEND: "true"
 kind: ConfigMap
 metadata:
   name: localai-env

helm/templates/services/localai/localai.deployment.yaml

@@ -4,7 +4,7 @@ metadata:
   name: localai
   namespace: {{ .Release.Namespace }}
 {{ tuple . "localai" | include "localai.labels" | indent 2 }}
-
+    intel.feature.node.kubernetes.io/gpu: "true"
 spec:
   replicas: 1
   selector:
@@ -13,6 +13,8 @@ spec:
   template:
     metadata:
 {{ tuple . "localai" | include "localai.labels" | indent 6 }}
+        intel.feature.node.kubernetes.io/gpu: "true"
+
     spec:
       containers:
         - name: localai
@@ -40,20 +42,22 @@ spec:
           volumeMounts:
             - name: localai-models-pv
               mountPath: /models
-            - name: localai-backends-pv
-              mountPath: /backends
+            # - name: localai-backends-pv
+            #   mountPath: /backends
           resources:
             limits:
               memory: 8Gi
               cpu: 4000m
+              gpu.intel.com/i915: 1
             requests:
               memory: 4Gi
               cpu: 1000m
+              gpu.intel.com/i915: 1
       restartPolicy: Always
       volumes:
         - name: localai-models-pv
           persistentVolumeClaim:
             claimName: localai-models-pv-claim
-        - name: localai-backends-pv
-          persistentVolumeClaim:
-            claimName: localai-backends-pv-claim
+        # - name: localai-backends-pv
+        #   persistentVolumeClaim:
+        #     claimName: localai-backends-pv-claim

packages/@liexp/backend/package.json

@@ -73,7 +73,7 @@
     "vitest-mock-extended": "^3.1.0"
   },
   "peerDependencies": {
-    "@langchain/core": "^0.3.19",
+    "@langchain/core": "^0.3.77",
     "@napi-rs/canvas": "^0.1.73",
     "debug": "^4.4.0",
     "effect": "^3.17.13",

packages/@liexp/backend/src/flows/ai/createEventFromDocuments.flow.ts

@@ -31,7 +31,7 @@ export const createEventFromDocuments = <
   return pipe(
     fp.RTE.Do,
     fp.RTE.bind("prompt", () => getCreateEventPromptPartial<C>(prompt, type)),
-    fp.RTE.bind("retriever", () => getStoreRetriever(documents)),
+    fp.RTE.bind("retriever", () => getStoreRetriever(documents, question)),
     fp.RTE.bind(
       "model",
       () => (ctx) =>
@@ -48,7 +48,7 @@ export const createEventFromDocuments = <
 
       return runRagChain<EventCommonProps, C>(
         {
-          context: retriever.pipe(formatDocumentsAsString),
+          context: () => formatDocumentsAsString(retriever),
         },
         prompt.pipe(model).pipe(parser),
         question,

packages/@liexp/backend/src/flows/ai/runRagChain.ts

@@ -39,30 +39,49 @@ export const runRagChain =
     }, toAPIError);
   };
 
-export const runRagChainStream =
+export const runRagChainStream = <
+  C extends LangchainContext & LoggerContext = LangchainContext & LoggerContext,
+>(
+  inputs: RunnableLike,
+  chain: RunnableLike,
+  question: string,
+  mode: "stream" | "invoke" = "invoke",
+): ReaderTaskEither<C, APIError, string> => {
+  return runRunnableSequence(
+    RunnableSequence.from([
+      { ...inputs, question: new RunnablePassthrough() },
+      chain,
+    ]),
+    question,
+    mode,
+  );
+};
+
+export const runRunnableSequence =
   <
     C extends LangchainContext & LoggerContext = LangchainContext &
       LoggerContext,
   >(
-    inputs: RunnableLike,
-    chain: RunnableLike,
+    inputs: RunnableSequence,
     question: string,
+    mode: "stream" | "invoke" = "stream",
   ): ReaderTaskEither<C, APIError, string> =>
   (ctx) => {
     return fp.TE.tryCatch(async () => {
-      const ragChain = RunnableSequence.from([
-        { ...inputs, question: new RunnablePassthrough() },
-        chain,
-      ]);
-
-      const stream = await ragChain.stream(question);
+      ctx.logger.debug.log("Running sequence in mode %s", mode);
 
       let output = "";
-      for await (const chunk of stream) {
-        output += chunk;
+      if (mode === "stream") {
+        const stream = await inputs.stream(question);
+
+        for await (const chunk of stream) {
+          output += chunk;
+        }
+      } else {
+        output = await inputs.invoke(question);
       }
 
-      ctx.logger.debug.log("RAG chain stream output %O", output);
+      ctx.logger.debug.log("Output %s", output);
 
       return output;
     }, toAPIError);

packages/@liexp/backend/src/flows/ai/storeRetriever.flow.ts

@@ -1,5 +1,4 @@
 import { type Document } from "@langchain/core/documents";
-import { type VectorStoreRetriever } from "@langchain/core/vectorstores";
 import { fp, pipe } from "@liexp/core/lib/fp/index.js";
 import {
   toAPIError,
@@ -14,16 +13,22 @@ import { type LoggerContext } from "../../context/logger.context.js";
 export const getStoreRetriever =
   <C extends LangchainContext & LoggerContext>(
     documents: Document[],
-  ): ReaderTaskEither<C, APIError, VectorStoreRetriever<MemoryVectorStore>> =>
+    question: string,
+  ): ReaderTaskEither<C, APIError, Document[]> =>
   (ctx) => {
     return pipe(
       fp.TE.tryCatch(async () => {
         const textSplitter = new RecursiveCharacterTextSplitter({
-          chunkSize: 2000,
-          chunkOverlap: 1000,
+          chunkSize: 1000,
+          chunkOverlap: 200,
         });
         const splits = await textSplitter.splitDocuments(documents);
 
+        ctx.logger.debug.log(
+          `Split documents into %d sub-documents.`,
+          splits.length,
+        );
+
         const vectorStore = await MemoryVectorStore.fromDocuments(
           splits,
           ctx.langchain.embeddings,
@@ -34,7 +39,11 @@ export const getStoreRetriever =
           documents.length,
         );
 
-        return vectorStore.asRetriever({ verbose: true });
+        const retriever = vectorStore.asRetriever({ verbose: true });
+
+        const retrieved = await retriever.invoke(question);
+
+        return retrieved;
       }, toAPIError),
     );
   };

packages/@liexp/backend/src/flows/ai/updateEventFromDocuments.flow.ts

@@ -30,13 +30,13 @@ export const updateEventFromDocuments = <
 ): ReaderTaskEither<C, APIError, EventCommonProps> => {
   return pipe(
     fp.RTE.Do,
-    fp.RTE.bind("prompt", () => getCreateEventPromptPartial<C>(prompt, type)),
-    fp.RTE.bind("retriever", () => getStoreRetriever(documents)),
+    fp.RTE.bind("prompt", () => getCreateEventPromptPartial(prompt, type)),
+    fp.RTE.bind("retriever", () => getStoreRetriever(documents, question)),
     fp.RTE.bind(
       "model",
       () => (ctx) =>
         fp.TE.right(
-          ctx.langchain.chat.bind({
+          ctx.langchain.chat.withConfig({
             response_format: {
               type: "json_object",
             },
@@ -48,7 +48,7 @@ export const updateEventFromDocuments = <
 
       return runRagChain<EventCommonProps, C>(
         {
-          context: retriever.pipe(formatDocumentsAsString),
+          context: () => formatDocumentsAsString(retriever),
         },
         prompt.pipe(model).pipe(parser),
         question,

packages/@liexp/backend/src/providers/ai/langchain.provider.ts

@@ -78,7 +78,7 @@ export interface LangchainProvider {
   chat: ChatOpenAI;
   embeddings: OpenAIEmbeddings;
   queryDocument: <Args extends { text: string; question?: string }>(
-    url: LangchainDocument[],
+    docs: LangchainDocument[],
     question: string,
     options?: { model?: AvailableModels; prompt?: PromptFn<Args> },
   ) => Promise<string>;
@@ -102,14 +102,15 @@ export const GetLangchainProvider = (
   const options = {
     ...opts,
   };
+
   const chat = new ChatOpenAI({
     model: chatModel,
     temperature: 0,
     apiKey: opts.apiKey,
     timeout: 60 * 30 * 1000, // 30 minutes
     maxConcurrency: 1,
     maxRetries: 2,
-    streamUsage: false,
+    streaming: true,
     ...opts.options?.chat,
     configuration: {
       baseURL: opts.baseURL,
@@ -121,6 +122,7 @@ export const GetLangchainProvider = (
 
   const embeddings = new OpenAIEmbeddings({
     model: embeddingsModel,
+    modelName: embeddingsModel,
     apiKey: opts.apiKey,
     timeout: 60 * 30 * 1000, // 30 minutes,
     ...opts.options?.embeddings,
@@ -141,20 +143,19 @@ export const GetLangchainProvider = (
     chat,
     embeddings,
     queryDocument: async (content, question, options) => {
-      const model =
-        options?.model ?? opts.models?.embeddings ?? "text-embedding-ada-002";
+      const model = options?.model ?? embeddingsModel;
 
-      const chatModel = options?.model ?? opts.models?.chat ?? "gpt-4o";
+      const chatModel2 = options?.model ?? chatModel;
 
       langchainLogger.info.log(
         "queryDocument use embedding model %s to query document with size %d using chat model %s",
         model,
         content.length,
-        chatModel,
+        chatModel2,
       );
 
       const chat = new ChatOpenAI({
-        model: chatModel,
+        model: chatModel2,
         temperature: 0,
         apiKey: opts.apiKey,
         configuration: {
@@ -168,6 +169,7 @@ export const GetLangchainProvider = (
         model,
         apiKey: opts.apiKey,
         timeout: 60 * 30 * 1000, // 30 minutes
+        encodingFormat: "base64",
         configuration: {
           baseURL: opts.baseURL,
           ...opts.options?.embeddings.configuration,
@@ -229,7 +231,7 @@ export const GetLangchainProvider = (
         configuration: {
           baseURL: opts.baseURL,
         },
-        streaming: true,
+        // streaming: true,
       });
 
       const textSplitter = new RecursiveCharacterTextSplitter({

packages/@liexp/backend/src/providers/fs/fs.provider.ts

@@ -71,7 +71,7 @@ export const GetFSClient = (ctx: GetFSClientContext): FSClient => {
   };
 
   const getObject: FSClient["getObject"] = (filePath) => {
-    fsLogger.debug.log("Getting object from path %s", filePath);
+    // fsLogger.debug.log("Getting object from path %s", filePath);
     return pipe(
       fp.IOE.tryCatch(
         () => ctx.client.readFileSync(filePath, "utf-8"),

packages/@liexp/ui/package.json

@@ -135,7 +135,8 @@
   "peerDependencies": {
     "axios": "^1",
     "effect": "^3",
-    "fp-ts": "^2"
+    "fp-ts": "^2",
+    "openai": "^5.20.2"
   },
   "packageManager": "[email protected]+sha512.34e538c329b5553014ca8e8f4535997f96180a1d0f614339357449935350d924e22f8614682191264ec33d1462ac21561aff97f6bb18065351c162c7e8f6de67"
 }