fix: unbreak document ingestion (docling to_formats) and OpenRouter chat model ids

apps/web/config/chat-models.yaml

@@ -17,27 +17,29 @@
 version: 1
 
 models:
-  # Default: Google Gemini. Reached at
-  # CHAT_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai,
-  # which is also where chat falls back when CHAT_BASE_URL is unset.
+  # Default: Google Gemini, reached through OpenRouter
+  # (CHAT_BASE_URL=https://openrouter.ai/api/v1).
   #
-  # Model ids here are the BARE ids Google's endpoint expects. The `google/`
-  # prefix belongs only to the `preset:` field, which names a bundled behavior
-  # entry — see packages/core/src/llm/presets.ts.
+  # Ids are sent to the endpoint verbatim, so they carry OpenRouter's `google/`
+  # vendor prefix. Talking to Google's own compatibility endpoint
+  # (https://generativelanguage.googleapis.com/v1beta/openai, also the fallback
+  # when CHAT_BASE_URL is unset) instead? Drop the prefix from `id:` — it
+  # expects the bare ids. The `preset:` field is unaffected either way: it names
+  # a bundled behavior entry, not a model on the endpoint.
   primary:
-    id: gemini-2.5-flash
+    id: google/gemini-2.5-flash
     preset: google/gemini-2.5-flash
 
   # The cheap tier. Its preset sets reasoning `default: none`, so routine work
   # does not pay for thinking tokens.
   fast:
-    id: gemini-2.5-flash-lite
+    id: google/gemini-2.5-flash-lite
     preset: google/gemini-2.5-flash-lite
 
   # The reasoning tier. 2.5 Pro always reasons, so its preset omits the `none`
   # level rather than claiming a capability the endpoint would reject.
   deep:
-    id: gemini-2.5-pro
+    id: google/gemini-2.5-pro
     preset: google/gemini-2.5-pro
 
   # Talking to something else instead? Any OpenAI-compatible endpoint works —

services/document-converter/src/docling.ts

@@ -1,6 +1,6 @@
 /**
  * docling-serve client — the consolidation of the old ocr-worker's
- * docling_runner.py. Same conversion parameters (to_formats=["md"], do_ocr,
+ * docling_runner.py. Same conversion parameters (to_formats=md, do_ocr,
  * do_table_structure, image_export_mode=placeholder); failures surface as
  * typed ServiceErrors instead of opaque 500s.
  */
@@ -25,7 +25,10 @@ export const doclingConvertFile: DoclingConvert = async (config, file, filename)
 
     const form = new FormData();
     form.append("files", new Blob([new Uint8Array(file)]), filename);
-    form.append("to_formats", '["md"]');
+    // One repeated form field per format. docling-serve validates each entry
+    // against its OutputFormat enum, so a JSON-encoded array ("[\"md\"]")
+    // arrives as a single bogus member and the request fails 422.
+    form.append("to_formats", "md");
     form.append("do_ocr", "true");
     form.append("do_table_structure", "true");
     form.append("image_export_mode", "placeholder");

services/document-converter/test/docling.test.ts

@@ -39,7 +39,7 @@ describe("doclingConvertFile outgoing request", () => {
         expect(form).toBeInstanceOf(FormData);
         expect(form!.get("md_page_break_placeholder")).toBe(PAGE_BREAK_PLACEHOLDER);
         // Pre-existing conversion parameters stay intact.
-        expect(form!.get("to_formats")).toBe('["md"]');
+        expect(form!.get("to_formats")).toBe("md");
         expect(form!.get("do_ocr")).toBe("true");
         expect(form!.get("do_table_structure")).toBe("true");
         expect(form!.get("image_export_mode")).toBe("placeholder");