Fix quality mode 500: pass ClaudeService instance to PdfConverter

marker's PdfConverter.__init__ eagerly resolves `default_llm_service`
(GoogleGeminiService) when `use_llm=True` *before* it applies any
config-dict override. So naming "marker.services.claude.ClaudeService"
in config["llm_service"] never took effect — Gemini's
verify_config_keys check fired first and 500'd on missing
gemini_api_key.

Build the ClaudeService instance ourselves and pass it via the
PdfConverter `llm_service=` constructor arg; that takes precedence
over the default and skips the Gemini path entirely.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

_project/CHANGELOG.md

@@ -8,6 +8,7 @@ Newest first. Format: `## YYYY-MM-DD — short title`, then bullets.
 
 ## Unreleased
 
+- **Fix:** Instantiate `ClaudeService` ourselves in `app/parser.py` and pass it to `PdfConverter(llm_service=…)` instead of naming the service via `config["llm_service"]`. Marker's `PdfConverter.__init__` eagerly resolves `default_llm_service` (Gemini) when `use_llm=True` *before* applying the config override, so the previous setup tripped Gemini's `gemini_api_key` check even though we never wanted Gemini. Surfaced as 500 from `/parse` in `quality` mode.
 - **Fix:** Pre-create `<site-packages>/static/fonts/` and chown to the `app` user. marker-pdf's `PdfConverter.__init__` calls `download_font()`, which writes there on first parse — without the chown the non-root container user got `PermissionError`, surfaced as 500 from `/parse`. Affected both `fast` and `quality` modes.
 - **Fix:** Redirect `~/.cache` to the bucket via `XDG_CACHE_HOME=/data/cache` and add `DATALAB_CACHE_DIR=/data/datalab` for explicit coverage. marker uses Vik's `datalab` loader for text-detection and OCR-error-detection models, which writes to `~/.cache/datalab/` by default. Without the redirect those models re-downloaded into the container's ephemeral overlay on every cold start (despite the bucket being mounted), adding seconds of avoidable latency. Also pinned `HF_HUB_CACHE` and `TRANSFORMERS_CACHE` explicitly under `/data/hf/`.
 

app/parser.py

@@ -16,6 +16,7 @@ import torch
 from marker.converters.pdf import PdfConverter
 from marker.models import create_model_dict
 from marker.output import text_from_rendered
+from marker.services.claude import ClaudeService
 
 log = logging.getLogger(__name__)
 
@@ -35,22 +36,31 @@ def device() -> str:
 def parse(pdf_path: str | Path, mode: Mode = "fast") -> tuple[str, int]:
     """Run marker on the PDF at `pdf_path`. Returns (markdown, page_count)."""
     config: dict[str, object] = {"output_format": "markdown"}
+    llm_service: ClaudeService | None = None
     if mode == "quality":
         api_key = os.getenv("ANTHROPIC_API_KEY")
         if not api_key:
             raise RuntimeError(
                 "parse_mode=quality requires ANTHROPIC_API_KEY to be set."
             )
-        config.update(
-            {
-                "use_llm": True,
-                "llm_service": "marker.services.claude.ClaudeService",
+        config["use_llm"] = True
+        # Pass the service instance directly. Marker's PdfConverter eagerly
+        # instantiates `default_llm_service` (Gemini) when `use_llm=True`,
+        # *before* applying any config-dict override — so naming the service
+        # in `config["llm_service"]` is too late and trips Gemini's config
+        # check. Supplying the instance via the constructor arg skips that.
+        llm_service = ClaudeService(
+            config={
                 "claude_api_key": api_key,
                 "claude_model_name": CLAUDE_MODEL,
             }
         )
 
-    converter = PdfConverter(artifact_dict=_models, config=config)
+    converter = PdfConverter(
+        artifact_dict=_models,
+        config=config,
+        llm_service=llm_service,
+    )
     rendered = converter(str(pdf_path))
     text, _, _ = text_from_rendered(rendered)
     return text, _page_count(rendered, pdf_path)