feat: add OpenAI-compatible embedding provider support

Add multi-provider embedding backend via CONTEXTPLUS_EMBED_PROVIDER env var:
- 'ollama' (default): existing Ollama backend, fully backward compatible
- 'openai': any OpenAI-compatible API (Gemini, OpenAI, Groq, etc.)

New env vars for OpenAI provider:
- CONTEXTPLUS_EMBED_PROVIDER=openai
- CONTEXTPLUS_OPENAI_API_KEY (or OPENAI_API_KEY)
- CONTEXTPLUS_OPENAI_BASE_URL (or OPENAI_BASE_URL)
- CONTEXTPLUS_OPENAI_EMBED_MODEL (or OPENAI_EMBED_MODEL)
- CONTEXTPLUS_OPENAI_CHAT_MODEL (or OPENAI_CHAT_MODEL)

Both embeddings.ts and semantic-navigate.ts patched.
Ollama import is now lazy (dynamic import) - no crash if Ollama not installed.

Author: Cenk Tekin

package-lock.json

@@ -1,7 +1,7 @@
-// Ollama-powered vector embedding engine with cosine similarity search
+// Multi-provider vector embedding engine with cosine similarity search
+// Supports Ollama (local) and OpenAI-compatible APIs (Gemini, OpenAI, etc.)
 // Indexes file headers and symbols, caches embeddings to disk for speed
 
-import { Ollama } from "ollama";
 import { readFile, writeFile, mkdir } from "fs/promises";
 import { join } from "path";
 
@@ -74,7 +74,11 @@ export interface EmbeddingCache {
   [path: string]: { hash: string; vector: number[] };
 }
 
+const EMBED_PROVIDER = (process.env.CONTEXTPLUS_EMBED_PROVIDER ?? "ollama").toLowerCase();
 const EMBED_MODEL = process.env.OLLAMA_EMBED_MODEL ?? "nomic-embed-text";
+const OPENAI_EMBED_MODEL = process.env.CONTEXTPLUS_OPENAI_EMBED_MODEL ?? process.env.OPENAI_EMBED_MODEL ?? "text-embedding-3-small";
+const OPENAI_API_KEY = process.env.CONTEXTPLUS_OPENAI_API_KEY ?? process.env.OPENAI_API_KEY ?? "";
+const OPENAI_BASE_URL = process.env.CONTEXTPLUS_OPENAI_BASE_URL ?? process.env.OPENAI_BASE_URL ?? "https://api.openai.com/v1";
 const CACHE_DIR = ".mcp_data";
 const CACHE_FILE = "embeddings-cache.json";
 const MIN_EMBED_BATCH_SIZE = 5;
@@ -87,7 +91,53 @@ const MIN_EMBED_CHUNK_CHARS = 256;
 const DEFAULT_EMBED_CHUNK_CHARS = 2000;
 const MAX_EMBED_CHUNK_CHARS = 8000;
 
-const ollama = new Ollama({ host: process.env.OLLAMA_HOST });
+type OllamaEmbedClient = { embed: (params: Record<string, unknown>) => Promise<{ embeddings: number[][] }> };
+let ollamaClient: OllamaEmbedClient | null = null;
+
+async function getOllamaClient(): Promise<OllamaEmbedClient> {
+  if (!ollamaClient) {
+    const { Ollama } = await import("ollama");
+    ollamaClient = new Ollama({ host: process.env.OLLAMA_HOST }) as unknown as OllamaEmbedClient;
+  }
+  return ollamaClient;
+}
+
+async function callOllamaEmbed(input: string[], signal: AbortSignal): Promise<number[][]> {
+  const client = await getOllamaClient();
+  const options = getEmbedRuntimeOptions();
+  const request: Record<string, unknown> = { model: EMBED_MODEL, input, signal };
+  if (options) request.options = options;
+  const response = await client.embed(request);
+  return response.embeddings;
+}
+
+async function callOpenAIEmbed(input: string[], signal: AbortSignal): Promise<number[][]> {
+  const url = `${OPENAI_BASE_URL.replace(/\/+$/, "")}/embeddings`;
+  const response = await fetch(url, {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/json",
+      "Authorization": `Bearer ${OPENAI_API_KEY}`,
+    },
+    body: JSON.stringify({ model: OPENAI_EMBED_MODEL, input }),
+    signal,
+  });
+
+  if (!response.ok) {
+    const body = await response.text().catch(() => "");
+    throw new Error(`OpenAI embed API error ${response.status}: ${body}`);
+  }
+
+  const data = await response.json() as { data: { embedding: number[] }[] };
+  return data.data.map((item) => item.embedding);
+}
+
+async function callProviderEmbed(input: string[], signal: AbortSignal): Promise<number[][]> {
+  if (EMBED_PROVIDER === "openai") {
+    return callOpenAIEmbed(input, signal);
+  }
+  return callOllamaEmbed(input, signal);
+}
 
 function toIntegerOr(value: string | undefined, fallback: number): number {
   if (!value) return fallback;
@@ -110,6 +160,7 @@ function toOptionalBoolean(value: string | undefined): boolean | undefined {
 }
 
 function getEmbedRuntimeOptions(): EmbedRuntimeOptions | undefined {
+  if (EMBED_PROVIDER === "openai") return undefined;
   const options: EmbedRuntimeOptions = {
     num_gpu: toOptionalInteger(process.env.CONTEXTPLUS_EMBED_NUM_GPU),
     main_gpu: toOptionalInteger(process.env.CONTEXTPLUS_EMBED_MAIN_GPU),
@@ -123,17 +174,6 @@ function getEmbedRuntimeOptions(): EmbedRuntimeOptions | undefined {
   return options;
 }
 
-function buildEmbedRequest(input: string[]): { model: string; input: string[]; options?: EmbedRuntimeOptions } {
-  const options = getEmbedRuntimeOptions();
-  return options ? { model: EMBED_MODEL, input, options } : { model: EMBED_MODEL, input };
-}
-
-async function embedWithTimeout(request: ReturnType<typeof buildEmbedRequest>): Promise<{ embeddings: number[][] }> {
-  const timeoutCtrl = AbortSignal.timeout(EMBED_TIMEOUT_MS);
-  const signal = AbortSignal.any([embedAbortController.signal, timeoutCtrl]);
-  return ollama.embed({ ...request, signal } as Parameters<typeof ollama.embed>[0]);
-}
-
 export function getEmbeddingBatchSize(): number {
   const requested = toIntegerOr(process.env.CONTEXTPLUS_EMBED_BATCH_SIZE, DEFAULT_EMBED_BATCH_SIZE);
   return Math.min(MAX_EMBED_BATCH_SIZE, Math.max(MIN_EMBED_BATCH_SIZE, requested));
@@ -152,7 +192,8 @@ function getErrorMessage(error: unknown): string {
 function isContextLengthError(error: unknown): boolean {
   const message = getErrorMessage(error).toLowerCase();
   return message.includes("input length exceeds context length")
-    || (message.includes("context") && message.includes("exceed"));
+    || (message.includes("context") && message.includes("exceed"))
+    || message.includes("maximum context length");
 }
 
 function shrinkEmbeddingInput(input: string): string {
@@ -167,9 +208,11 @@ async function embedSingleAdaptive(input: string): Promise<number[]> {
 
   for (let attempt = 0; attempt <= MAX_SINGLE_INPUT_RETRIES; attempt++) {
     try {
-      const response = await embedWithTimeout(buildEmbedRequest([candidate]));
-      if (!response.embeddings[0]) throw new Error("Missing embedding vector in Ollama response");
-      return response.embeddings[0];
+      const timeoutCtrl = AbortSignal.timeout(EMBED_TIMEOUT_MS);
+      const signal = AbortSignal.any([embedAbortController.signal, timeoutCtrl]);
+      const embeddings = await callProviderEmbed([candidate], signal);
+      if (!embeddings[0]) throw new Error("Missing embedding vector in response");
+      return embeddings[0];
     } catch (error) {
       if (!isContextLengthError(error)) throw error;
       const nextCandidate = shrinkEmbeddingInput(candidate);
@@ -183,11 +226,13 @@ async function embedSingleAdaptive(input: string): Promise<number[]> {
 
 async function embedBatchAdaptive(batch: string[]): Promise<number[][]> {
   try {
-    const response = await embedWithTimeout(buildEmbedRequest(batch));
-    if (response.embeddings.length !== batch.length) {
-      throw new Error(`Embedding response size mismatch: expected ${batch.length}, got ${response.embeddings.length}`);
+    const timeoutCtrl = AbortSignal.timeout(EMBED_TIMEOUT_MS);
+    const signal = AbortSignal.any([embedAbortController.signal, timeoutCtrl]);
+    const embeddings = await callProviderEmbed(batch, signal);
+    if (embeddings.length !== batch.length) {
+      throw new Error(`Embedding response size mismatch: expected ${batch.length}, got ${embeddings.length}`);
     }
-    return response.embeddings;
+    return embeddings;
   } catch (error) {
     if (!isContextLengthError(error)) throw error;
     if (batch.length === 1) {

src/core/embeddings.ts

@@ -1,7 +1,6 @@
-// Semantic project navigator using spectral clustering and Ollama labeling
+// Semantic project navigator using spectral clustering and provider-agnostic labeling
 // Browse codebase by meaning: embeds files, clusters vectors, generates labels
 
-import { Ollama } from "ollama";
 import { walkDirectory } from "../core/walker.js";
 import { analyzeFile, flattenSymbols, isSupportedFile } from "../core/parser.js";
 import { fetchEmbedding } from "../core/embeddings.js";
@@ -29,8 +28,11 @@ interface ClusterNode {
   children: ClusterNode[];
 }
 
-const EMBED_MODEL = process.env.OLLAMA_EMBED_MODEL ?? "nomic-embed-text";
+const EMBED_PROVIDER = (process.env.CONTEXTPLUS_EMBED_PROVIDER ?? "ollama").toLowerCase();
 const CHAT_MODEL = process.env.OLLAMA_CHAT_MODEL ?? "llama3.2";
+const OPENAI_CHAT_MODEL = process.env.CONTEXTPLUS_OPENAI_CHAT_MODEL ?? process.env.OPENAI_CHAT_MODEL ?? "gpt-4o-mini";
+const OPENAI_API_KEY = process.env.CONTEXTPLUS_OPENAI_API_KEY ?? process.env.OPENAI_API_KEY ?? "";
+const OPENAI_BASE_URL = process.env.CONTEXTPLUS_OPENAI_BASE_URL ?? process.env.OPENAI_BASE_URL ?? "https://api.openai.com/v1";
 const MAX_FILES_PER_LEAF = 20;
 const NON_CODE_NAVIGATE_EXTENSIONS = new Set([
   ".json",
@@ -46,7 +48,16 @@ const NON_CODE_NAVIGATE_EXTENSIONS = new Set([
   ".env",
 ]);
 
-const ollama = new Ollama({ host: process.env.OLLAMA_HOST });
+type OllamaChatClient = { chat: (params: Record<string, unknown>) => Promise<{ message: { content: string } }> };
+let ollamaClient: OllamaChatClient | null = null;
+
+async function getOllamaClient(): Promise<OllamaChatClient> {
+  if (!ollamaClient) {
+    const { Ollama } = await import("ollama");
+    ollamaClient = new Ollama({ host: process.env.OLLAMA_HOST }) as unknown as OllamaChatClient;
+  }
+  return ollamaClient;
+}
 
 async function fetchEmbeddings(inputs: string[]): Promise<number[][]> {
   return fetchEmbedding(inputs);
@@ -57,7 +68,32 @@ function isNavigableSourceCandidate(filePath: string): boolean {
 }
 
 async function chatCompletion(prompt: string): Promise<string> {
-  const response = await ollama.chat({
+  if (EMBED_PROVIDER === "openai") {
+    const url = `${OPENAI_BASE_URL.replace(/\/+$/, "")}/chat/completions`;
+    const response = await fetch(url, {
+      method: "POST",
+      headers: {
+        "Content-Type": "application/json",
+        "Authorization": `Bearer ${OPENAI_API_KEY}`,
+      },
+      body: JSON.stringify({
+        model: OPENAI_CHAT_MODEL,
+        messages: [{ role: "user", content: prompt }],
+        stream: false,
+      }),
+    });
+
+    if (!response.ok) {
+      const body = await response.text().catch(() => "");
+      throw new Error(`OpenAI chat API error ${response.status}: ${body}`);
+    }
+
+    const data = await response.json() as { choices: { message: { content: string } }[] };
+    return data.choices[0]?.message?.content ?? "";
+  }
+
+  const client = await getOllamaClient();
+  const response = await client.chat({
     model: CHAT_MODEL,
     messages: [{ role: "user", content: prompt }],
     stream: false,
@@ -123,7 +159,7 @@ async function labelSiblingClusters(clusters: { files: FileInfo[]; pathPattern:
 
   const prompt = `You are labeling clusters of code files. For each cluster below, produce EXACTLY one JSON array of objects, each with:
 - "overarchingTheme": a sentence about the cluster's theme
-- "distinguishingFeature": what makes this cluster unique vs siblings  
+- "distinguishingFeature": what makes this cluster unique vs siblings
 - "label": EXACTLY 2 words describing the cluster
 
 ${clusterDescriptions.join("\n\n")}
@@ -256,7 +292,10 @@ export async function semanticNavigate(options: SemanticNavigateOptions): Promis
     vectors = embedded.vectors;
     skippedForEmbedding = embedded.skipped;
   } catch (err) {
-    return `Ollama not available for embeddings: ${err instanceof Error ? err.message : String(err)}\nMake sure Ollama is running or signed in (ollama signin) with model ${EMBED_MODEL}.`;
+    const providerHint = EMBED_PROVIDER === "openai"
+      ? `Check CONTEXTPLUS_OPENAI_API_KEY and CONTEXTPLUS_OPENAI_BASE_URL.`
+      : `Make sure Ollama is running with model ${CHAT_MODEL}.`;
+    return `Embedding provider (${EMBED_PROVIDER}) not available: ${err instanceof Error ? err.message : String(err)}\n${providerHint}`;
   }
 
   if (embeddableFiles.length === 0) return "No embeddable source files found in the project.";