@heripo/document-processor 0.1.32 → 0.1.33

package/dist/index.js

@@ -86,6 +86,8 @@ function detectProvider(model) {
   if (providerId.includes("google")) return "google";
   if (providerId.includes("anthropic")) return "anthropic";
   if (providerId.includes("together")) return "togetherai";
+  if (providerId.includes("ollama")) return "ollama";
+  if (providerId.includes("lmstudio")) return "lmstudio";
   return "unknown";
 }
 var LLMCaller = class {
@@ -173,14 +175,18 @@ var LLMCaller = class {
    *
    * Strategy per provider:
    * - OpenAI / Anthropic / Google Gemini: `Output.object()` with schema retry
-   * - Together AI / unknown: forced tool call pattern
+   * - Together AI / Ollama / unknown: forced tool call pattern
+   *
+   * Ollama uses tool-call because MLX backends ignore the `format` grammar
+   * that `Output.object()` relies on, and gguf's grammar is unstable once
+   * thinking is disabled. Applies to both text and vision paths.
    *
    * Retries up to MAX_STRUCTURED_OUTPUT_RETRIES times on NoObjectGeneratedError
    * (schema mismatch), re-throwing the last error if all attempts fail.
    */
   static async generateStructuredOutput(model, schema, promptParams) {
     const providerType = detectProvider(model);
-    if (providerType === "togetherai" || providerType === "unknown") {
+    if (providerType === "togetherai" || providerType === "ollama" || providerType === "unknown") {
       return this.generateViaToolCall(model, schema, promptParams);
     }
     let lastError;