feat: add Gemini embeddings support + Letta compatibility fixes

Cherry-picked from decolua/9router#148 (author: xuandung38 / Hồ Xuân Dũng <me@hxd.vn>)

- Add Google AI (Gemini) embeddings support for /v1/embeddings endpoint
- Add Gemini embedding models: gemini-embedding-001, text-embedding-005, text-embedding-004
- Inject missing object/created fields for Letta and strict OpenAI clients
- Strip Azure-specific fields (prompt_filter_results, content_filter_results) from responses
- Fix Dockerfile: copy open-sse directory into Docker runner stage

Skipped: whitelist message field stripping (commit 3/7/8) — too aggressive for all providers
Skipped: default stream=false change (commit 9) — behavior change needs further review
Co-authored-by: Cursor <cursoragent@cursor.com>

Dockerfile

@@ -22,6 +22,7 @@ RUN mkdir -p /app/data
 COPY --from=builder /app/public ./public
 COPY --from=builder /app/.next/static ./.next/static
 COPY --from=builder /app/.next/standalone ./
+COPY --from=builder /app/open-sse ./open-sse
 
 EXPOSE 20128
 

open-sse/config/providerModels.js

@@ -135,6 +135,10 @@ export const PROVIDER_MODELS = {
     { id: "gemini-2.5-pro", name: "Gemini 2.5 Pro" },
     { id: "gemini-2.5-flash", name: "Gemini 2.5 Flash" },
     { id: "gemini-2.5-flash-lite", name: "Gemini 2.5 Flash Lite" },
+    // Embedding models
+    { id: "gemini-embedding-001", name: "Gemini Embedding 001", type: "embedding" },
+    { id: "text-embedding-005", name: "Text Embedding 005", type: "embedding" },
+    { id: "text-embedding-004", name: "Text Embedding 004 (Legacy)", type: "embedding" },
   ],
   openrouter: [
     { id: "auto", name: "Auto (Best Available)" },

open-sse/handlers/chatCore.js

@@ -760,6 +760,22 @@ export async function handleChatCore({ body, modelInfo, credentials, log, onCred
       ? translateNonStreamingResponse(responseBody, targetFormat, sourceFormat)
       : responseBody;
 
+    // Ensure OpenAI-required fields are present (needed for Letta and other strict clients)
+    if (!translatedResponse.object) translatedResponse.object = "chat.completion";
+    if (!translatedResponse.created) translatedResponse.created = Math.floor(Date.now() / 1000);
+
+    // Strip Azure-specific non-standard fields
+    if (translatedResponse.prompt_filter_results !== undefined) {
+      delete translatedResponse.prompt_filter_results;
+    }
+    if (translatedResponse?.choices) {
+      for (const choice of translatedResponse.choices) {
+        if (choice.content_filter_results !== undefined) {
+          delete choice.content_filter_results;
+        }
+      }
+    }
+
     // Add buffer and filter usage for client (to prevent CLI context errors)
     if (translatedResponse?.usage) {
       const buffered = addBufferToUsage(translatedResponse.usage);

open-sse/handlers/embeddingsCore.js

@@ -4,16 +4,49 @@ import { HTTP_STATUS } from "../config/constants.js";
 import { getExecutor } from "../executors/index.js";
 import { refreshWithRetry } from "../services/tokenRefresh.js";
 
+// Google AI (Gemini) provider aliases / identifiers
+const GEMINI_PROVIDERS = new Set(["gemini", "google_ai_studio"]);
+
+/**
+ * Check whether a provider targets the Google AI (Gemini) embeddings API.
+ * @param {string} provider
+ */
+function isGeminiProvider(provider) {
+  return GEMINI_PROVIDERS.has(provider);
+}
+
 /**
  * Build the embeddings request body for the target provider.
- * Most OpenAI-compatible providers accept the same format.
+ *
- * For providers that don't support embeddings natively (chat-only), we return an error.
+ * - OpenAI / openai-compatible / openrouter: standard { model, input } format.
+ * - Google AI (Gemini): different format per API spec.
+ *   - Single input  → embedContent  body: { model, content: { parts: [{ text }] } }
+ *   - Batch input   → batchEmbedContents body: { requests: [{ model, content: { parts: [{ text }] } }] }
  */
-function buildEmbeddingsBody(model, input, encodingFormat) {
+function buildEmbeddingsBody(provider, model, input, encodingFormat) {
-  const body = {
+  if (isGeminiProvider(provider)) {
-    model,
+    // Normalize model name: Gemini API expects "models/<model>" prefix
-    input
+    const geminiModel = model.startsWith("models/") ? model : `models/${model}`;
+
+    if (Array.isArray(input)) {
+      // Batch request
+      return {
+        requests: input.map((text) => ({
+          model: geminiModel,
+          content: { parts: [{ text: String(text) }] }
+        }))
       };
+    } else {
+      // Single request
+      return {
+        model: geminiModel,
+        content: { parts: [{ text: String(input) }] }
+      };
+    }
+  }
+
+  // Default: OpenAI format
+  const body = { model, input };
   if (encodingFormat) {
     body.encoding_format = encodingFormat;
   }
@@ -22,8 +55,24 @@ function buildEmbeddingsBody(model, input, encodingFormat) {
 
 /**
  * Build the URL for the embeddings endpoint based on the provider.
+ * @param {string} provider
+ * @param {string} model
+ * @param {object} credentials
+ * @param {string|string[]} input - used to select single vs batch endpoint for Gemini
  */
-function buildEmbeddingsUrl(provider, credentials) {
+function buildEmbeddingsUrl(provider, model, credentials, input) {
+  if (isGeminiProvider(provider)) {
+    const apiKey = credentials.apiKey || credentials.accessToken;
+    // Normalize model name for URL path
+    const modelPath = model.startsWith("models/") ? model : `models/${model}`;
+
+    if (Array.isArray(input)) {
+      // batchEmbedContents for array input (keeps response format consistent even for length=1)
+      return `https://generativelanguage.googleapis.com/v1beta/${modelPath}:batchEmbedContents?key=${encodeURIComponent(apiKey)}`;
+    }
+    return `https://generativelanguage.googleapis.com/v1beta/${modelPath}:embedContent?key=${encodeURIComponent(apiKey)}`;
+  }
+
   switch (provider) {
     case "openai":
       return "https://api.openai.com/v1/embeddings";
@@ -46,6 +95,11 @@ function buildEmbeddingsUrl(provider, credentials) {
 function buildEmbeddingsHeaders(provider, credentials) {
   const headers = { "Content-Type": "application/json" };
 
+  if (isGeminiProvider(provider)) {
+    // Gemini API uses API key as query param — no Authorization header needed
+    return headers;
+  }
+
   switch (provider) {
     case "openai":
     case "openrouter":
@@ -56,11 +110,7 @@ function buildEmbeddingsHeaders(provider, credentials) {
       }
       break;
     default:
-      if (provider?.startsWith?.("openai-compatible-")) {
       headers["Authorization"] = `Bearer ${credentials.apiKey || credentials.accessToken}`;
-      } else {
-        headers["Authorization"] = `Bearer ${credentials.apiKey || credentials.accessToken}`;
-      }
   }
 
   return headers;
@@ -68,14 +118,52 @@ function buildEmbeddingsHeaders(provider, credentials) {
 
 /**
  * Normalize the embeddings response to OpenAI format.
- * Most OpenAI-compatible providers already return this format.
+ *
+ * Gemini single response:
+ *   { embedding: { values: [0.1, 0.2, ...] } }
+ *
+ * Gemini batch response:
+ *   { embeddings: [{ values: [...] }, ...] }
+ *
+ * Target OpenAI format:
+ *   { object: "list", data: [{ object: "embedding", index: 0, embedding: [...] }], model, usage: {...} }
  */
-function normalizeEmbeddingsResponse(responseBody, model) {
+function normalizeEmbeddingsResponse(responseBody, model, provider) {
   // Already in OpenAI format
   if (responseBody.object === "list" && Array.isArray(responseBody.data)) {
     return responseBody;
   }
 
+  if (isGeminiProvider(provider)) {
+    let embeddingItems = [];
+
+    if (Array.isArray(responseBody.embeddings)) {
+      // Batch response
+      embeddingItems = responseBody.embeddings.map((emb, idx) => ({
+        object: "embedding",
+        index: idx,
+        embedding: emb.values || []
+      }));
+    } else if (responseBody.embedding?.values) {
+      // Single response
+      embeddingItems = [{
+        object: "embedding",
+        index: 0,
+        embedding: responseBody.embedding.values
+      }];
+    }
+
+    return {
+      object: "list",
+      data: embeddingItems,
+      model,
+      usage: {
+        prompt_tokens: 0,
+        total_tokens: 0
+      }
+    };
+  }
+
   // Try to handle alternate formats gracefully
   return responseBody;
 }
@@ -114,16 +202,16 @@ export async function handleEmbeddingsCore({
   const encodingFormat = body.encoding_format || "float";
 
   // Determine embeddings URL
-  const url = buildEmbeddingsUrl(provider, credentials);
+  const url = buildEmbeddingsUrl(provider, model, credentials, input);
   if (!url) {
     return createErrorResult(
       HTTP_STATUS.BAD_REQUEST,
-      `Provider '${provider}' does not support embeddings. Use openai, openrouter, or an openai-compatible provider.`
+      `Provider '${provider}' does not support embeddings. Use openai, openrouter, gemini, or an openai-compatible provider.`
     );
   }
 
   const headers = buildEmbeddingsHeaders(provider, credentials);
-  const requestBody = buildEmbeddingsBody(model, input, encodingFormat);
+  const requestBody = buildEmbeddingsBody(provider, model, input, encodingFormat);
 
   log?.debug?.("EMBEDDINGS", `${provider.toUpperCase()} | ${model} | input_type=${Array.isArray(input) ? `array[${input.length}]` : "string"}`);
 
@@ -162,7 +250,12 @@ export async function handleEmbeddingsCore({
       // Retry with refreshed credentials
       try {
         const retryHeaders = buildEmbeddingsHeaders(provider, credentials);
-        providerResponse = await fetch(url, {
+        // Rebuild URL for Gemini since API key is embedded in query param
+        const retryUrl = isGeminiProvider(provider)
+          ? buildEmbeddingsUrl(provider, model, credentials, input)
+          : url;
+
+        providerResponse = await fetch(retryUrl, {
           method: "POST",
           headers: retryHeaders,
           body: JSON.stringify(requestBody)
@@ -193,7 +286,7 @@ export async function handleEmbeddingsCore({
     await onRequestSuccess();
   }
 
-  const normalized = normalizeEmbeddingsResponse(responseBody, model);
+  const normalized = normalizeEmbeddingsResponse(responseBody, model, provider);
 
   log?.debug?.("EMBEDDINGS", `Success | usage=${JSON.stringify(normalized.usage || {})}`);
 

open-sse/utils/stream.js

@@ -82,6 +82,25 @@ export function createSSEStream(options = {}) {
 
               const idFixed = fixInvalidId(parsed);
 
+              // Ensure OpenAI-required fields are present on streaming chunks (Letta compat)
+              let fieldsInjected = false;
+              if (!parsed.object) { parsed.object = "chat.completion.chunk"; fieldsInjected = true; }
+              if (!parsed.created) { parsed.created = Math.floor(Date.now() / 1000); fieldsInjected = true; }
+
+              // Strip Azure-specific non-standard fields from streaming chunks
+              if (parsed.prompt_filter_results !== undefined) {
+                delete parsed.prompt_filter_results;
+                fieldsInjected = true;
+              }
+              if (parsed?.choices) {
+                for (const choice of parsed.choices) {
+                  if (choice.content_filter_results !== undefined) {
+                    delete choice.content_filter_results;
+                    fieldsInjected = true;
+                  }
+                }
+              }
+
               if (!hasValuableContent(parsed, FORMATS.OPENAI)) {
                 continue;
               }
@@ -115,7 +134,7 @@ export function createSSEStream(options = {}) {
                 parsed.usage = filterUsageForFormat(buffered, FORMATS.OPENAI);
                 output = `data: ${JSON.stringify(parsed)}\n`;
                 injectedUsage = true;
-              } else if (idFixed) {
+              } else if (idFixed || fieldsInjected) {
                 output = `data: ${JSON.stringify(parsed)}\n`;
                 injectedUsage = true;
               }