| Environment:
- XWiki on Debian 13 / Tomcat 11
- LLM Application (BETA) 0.9 + Index for the LLM Application 0.9
- Backend: Ollama 0.34.0 (OpenAI-compatible /v1 API) behind a Caddy
reverse proxy with flush_interval -1 (no buffering)
- Model: gemma3:12b, RAG context enabled (8 chunks, ~2000 chars each)
- "Can Stream" enabled on the server configuration
Steps to reproduce:
1. Ask a question that produces a long streamed response (RAG context
included, total response body roughly 12-17 KB)
2. Intermittently, the chat shows:
"An error occurred: Unterminated string in JSON at position 13303
(line 1 column 13304)" Observed on multiple occasions at different positions (12363, 13303,
17438), always mid-string, always on longer responses. Short responses
stream fine. Workaround: unchecking "Can Stream" on the server configuration makes
the problem disappear entirely, so the response content itself is valid.
This suggests the client-side parser fails when a JSON string is split
across SSE chunk boundaries. |
