refactor embeddings

packages/adapter-postgres/seed.sql

@@ -1,3 +1,9 @@
-INSERT INTO public.accounts (id, name, email, avatarUrl, details) VALUES ('00000000-0000-0000-0000-000000000000', 'Default Agent', 'default@agent.com', '', '{}');
-INSERT INTO public.rooms (id) VALUES ('00000000-0000-0000-0000-000000000000');
-INSERT INTO public.participants (userId, roomId) VALUES ('00000000-0000-0000-0000-000000000000', '00000000-0000-0000-0000-000000000000');
+
+INSERT INTO public.accounts (id, name, email, "avatarUrl", details)
+VALUES ('00000000-0000-0000-0000-000000000000', 'Default Agent', 'default@agent.com', '', '{}'::jsonb);
+
+INSERT INTO public.rooms (id)
+VALUES ('00000000-0000-0000-0000-000000000000');
+
+INSERT INTO public.participants (id, "userId", "roomId")
+VALUES ('00000000-0000-0000-0000-000000000001', '00000000-0000-0000-0000-000000000000', '00000000-0000-0000-0000-000000000000');

packages/core/src/embedding.ts

@@ -22,105 +22,106 @@ function getRootPath() {
     return path.resolve(__dirname, "..");
 }
 
-/**
- * Send a message to the OpenAI API for embedding.
- * @param input The input to be embedded.
- * @returns The embedding of the input.
- */
-export async function embed(runtime: IAgentRuntime, input: string) {
-    // get the charcter, and handle by model type
-    const modelProvider = models[runtime.character.modelProvider];
-    const embeddingModel = modelProvider.model.embedding;
-
-    if (
-        runtime.character.modelProvider !== ModelProviderName.OPENAI &&
-        runtime.character.modelProvider !== ModelProviderName.OLLAMA &&
-        !settings.USE_OPENAI_EMBEDDING
-    ) {
-
-        // make sure to trim tokens to 8192
-        const cacheDir = getRootPath() + "/cache/";
-
-        // if the cache directory doesn't exist, create it
-        if (!fs.existsSync(cacheDir)) {
-            fs.mkdirSync(cacheDir, { recursive: true });
-        }
-
-        const embeddingModel = await FlagEmbedding.init({
-            cacheDir: cacheDir
-        });
-
-        const trimmedInput = trimTokens(input, 8000, "gpt-4o-mini");
-
-        const embedding: number[] = await embeddingModel.queryEmbed(trimmedInput);
-        console.log("Embedding dimensions: ", embedding.length);
-        return embedding;
-
-        // commented out the text generation service that uses llama
-        // const service = runtime.getService<ITextGenerationService>(
-        //     ServiceType.TEXT_GENERATION
-        // );
-
-        // const instance = service?.getInstance();
-
-        // if (instance) {
-        //     return await instance.getEmbeddingResponse(input);
-        // }
-    }
-
-    // TODO: Fix retrieveCachedEmbedding
-    // Check if we already have the embedding in the lore
-    const cachedEmbedding = await retrieveCachedEmbedding(runtime, input);
-    if (cachedEmbedding) {
-        return cachedEmbedding;
-    }
+interface EmbeddingOptions {
+    model: string;
+    endpoint: string;
+    apiKey?: string;
+    length?: number;
+    isOllama?: boolean;
+}
 
+async function getRemoteEmbedding(input: string, options: EmbeddingOptions): Promise<number[]> {
     const requestOptions = {
         method: "POST",
         headers: {
             "Content-Type": "application/json",
-            // TODO: make this not hardcoded
-            // TODO: make this not hardcoded
-            ...((runtime.modelProvider !== ModelProviderName.OLLAMA || settings.USE_OPENAI_EMBEDDING) ? {
-                Authorization: `Bearer ${runtime.token}`,
-                } : {}),
+            ...(options.apiKey ? {
+                Authorization: `Bearer ${options.apiKey}`,
+            } : {}),
         },
         body: JSON.stringify({
             input,
-            model: embeddingModel,
-            length: 384, // we are squashing dimensions to 768 for openai, even thought the model supports 1536
-            // -- this is ok for matryoshka embeddings but longterm, we might want to support 1536
+            model: options.model,
+            length: options.length || 384,
         }),
     };
+
     try {
         const response = await fetch(
-            // TODO: make this not hardcoded
-            `${runtime.character.modelEndpointOverride || modelProvider.endpoint}${(runtime.character.modelProvider === ModelProviderName.OLLAMA && !settings.USE_OPENAI_EMBEDDING) ? "/v1" : ""}/embeddings`,
+            `${options.endpoint}${options.isOllama ? "/v1" : ""}/embeddings`,
             requestOptions
         );
 
         if (!response.ok) {
             throw new Error(
-                "OpenAI API Error: " +
-                    response.status +
-                    " " +
-                    response.statusText
+                "Embedding API Error: " +
+                response.status +
+                " " +
+                response.statusText
             );
         }
 
-        interface OpenAIEmbeddingResponse {
+        interface EmbeddingResponse {
             data: Array<{ embedding: number[] }>;
         }
 
-        const data: OpenAIEmbeddingResponse = await response.json();
-
+        const data: EmbeddingResponse = await response.json();
         return data?.data?.[0].embedding;
     } catch (e) {
         console.error(e);
         throw e;
     }
 }
 
+/**
+ * Send a message to the OpenAI API for embedding.
+ * @param input The input to be embedded.
+ * @returns The embedding of the input.
+ */
+export async function embed(runtime: IAgentRuntime, input: string) {
+    const modelProvider = models[runtime.character.modelProvider];
+    const embeddingModel = modelProvider.model.embedding;
+
+    // Try local embedding first
+    if (
+        runtime.character.modelProvider !== ModelProviderName.OPENAI &&
+        runtime.character.modelProvider !== ModelProviderName.OLLAMA &&
+        !settings.USE_OPENAI_EMBEDDING
+    ) {
+        return await getLocalEmbedding(input);
+    }
+
+    // Check cache
+    const cachedEmbedding = await retrieveCachedEmbedding(runtime, input);
+    if (cachedEmbedding) {
+        return cachedEmbedding;
+    }
+
+    // Get remote embedding
+    return await getRemoteEmbedding(input, {
+        model: embeddingModel,
+        endpoint: runtime.character.modelEndpointOverride || modelProvider.endpoint,
+        apiKey: runtime.token,
+        isOllama: runtime.character.modelProvider === ModelProviderName.OLLAMA && !settings.USE_OPENAI_EMBEDDING
+    });
+}
+
+async function getLocalEmbedding(input: string): Promise<number[]> {
+    const cacheDir = getRootPath() + "/cache/";
+    if (!fs.existsSync(cacheDir)) {
+        fs.mkdirSync(cacheDir, { recursive: true });
+    }
+
+    const embeddingModel = await FlagEmbedding.init({
+        cacheDir: cacheDir
+    });
+
+    const trimmedInput = trimTokens(input, 8000, "gpt-4o-mini");
+    const embedding = await embeddingModel.queryEmbed(trimmedInput);
+    console.log("Embedding dimensions: ", embedding.length);
+    return embedding;
+}
+
 export async function retrieveCachedEmbedding(
     runtime: IAgentRuntime,
     input: string
@@ -129,11 +130,12 @@ export async function retrieveCachedEmbedding(
         console.log("No input to retrieve cached embedding for");
         return null;
     }
-
+    
     const similaritySearchResult =
         await runtime.messageManager.getCachedEmbeddings(input);
     if (similaritySearchResult.length > 0) {
         return similaritySearchResult[0].embedding;
     }
     return null;
 }
+