litellm.exceptions.APIConnectionError && No content chunks

[alrex66608]

q1 :使用私有化部署，发现ollama似乎兼容错误

q2 :文件识别失败，提示 No content chunks (text or image) could be extracted from the document
尽管我在 morphik.toml 中新增配置了依旧报错

[ingestion]
pdf_backend = "pymupdf"
enable_multimodal_parsing = false
min_chunk_length = 20

2026-01-12 09:18:01,001 - ERROR - Error processing ingestion job for file 测试PDF识别.pdf: No content chunks (text or image) could be extracted from the document
2026-01-12 09:18:01,003 - ERROR - Traceback (most recent call last):
File "/app/core/workers/ingestion_worker.py", line 783, in process_ingestion_job
raise ValueError("No content chunks (text or image) could be extracted from the document")
ValueError: No content chunks (text or image) could be extracted from the document

[BloggerBust]

Hey! Just checking in regarding the diagnostics PRs (#381 and #383) tied to this issue.

They improve error visibility around ingestion and LiteLLM completion failures and add unit tests for the cases discussed here.

I am happy to adjust the approach or split things up differently if that would make review easier.

[JiwaniZakir]

Two separate issues here. For the Ollama APIConnectionError, this is typically caused by a base URL mismatch — litellm expects Ollama to be reachable at the configured endpoint (usually http://localhost:11434), but in a containerized/private deployment the service name or port differs; check that api_base in your model config matches the actual reachable address from within the container network.

For the "No content chunks" error, the pymupdf backend is being set but the extraction in ingestion_worker.py around line 783 is still returning empty — this often happens when the PDF is scanned/image-only and enable_multimodal_parsing = false prevents image chunk extraction, leaving nothing to index. Try setting enable_multimodal_parsing = true (and ensuring a vision-capable model is configured) or running the PDF through OCR preprocessing before ingestion, since PyMuPDF alone cannot extract text from image-based PDFs.