mirror of
https://github.com/discourse/discourse.git
synced 2026-08-04 10:39:43 +08:00
c10cce7b72Follow-up to #41716, found by live-testing the new parsing chain against locally configured vLLM models (Jolteon/CDCK MoM, DeepSeek V4 Flash, Qwen3.6 35B). ## The bug DeepSeek V4 Flash reproducibly prefixes its structured output with a stray brace-quote, so the streamed document looks like: ``` {"{"output":"Olá a todos! ..."} ``` The complete, correct translation is in there, but every parser in the best-effort chain rejects the document (the stray prefix parses as an object key with no value), and `extract_key` returned an empty string — so translations from this model came back blank. ## The fix Add a recovery attempt to the chain, between the control-character re-escape and the lenient parser: locate the schema key in the text and re-parse (truncation-tolerant) from the object brace immediately preceding it. This is generic for any "garbage before the real object" response, not just the doubled brace-quote observed here. ## Verification Live runs of the real `PostRawTranslator` pipeline (hazard-rich source: emoji, inline code with `<`/`&`, code fences, blockquote, quoted strings) against the three vLLM models: | model | before | after | |---|---|---| | Jolteon (CDCK/MoM) | complete, clean stream | complete, clean stream | | Qwen3.6 35B A3B | complete, clean stream | complete, clean stream | | DeepSeek V4 Flash | **blank translation** | complete translation via fallback, warning logged | All structure checks pass on the recovered output: no control characters, no literal escape artifacts, angle brackets/query strings/code fences/blockquote/emoji intact. Plus a regression spec with the exact observed shape. Parser + structured output suites: 111 examples, 0 failures. |
||
|---|---|---|
| .. | ||
| diff_utils | ||
| research | ||
| ai_staff_action_logger.rb | ||
| best_effort_json_parser.rb | ||
| dns_srv.rb | ||
| image_to_text.rb | ||
| pdf_to_text.rb | ||
| search.rb | ||
