Skip to content

Magazine extraction fails silently: LLM response JSON not parsed, no diagnostic logging

Description

When running metaextractor extract on structured OCR JSON arrays (e.g., supermama.json), the magazine metadata pipeline fails with:

Magazine extraction prompt did not produce valid JSON

Three root causes:

  1. Data preparation bug: process_text_file() only handled dict JSON inputs. JSON arrays (lists of OCR blocks) fell through to the fallback path, which passed the raw JSON string (including all structural characters {, }, [, ]) as "OCR text" to the LLM. This garbled input confused the model.

  2. No diagnostic logging: When parse_json_object() failed, it silently returned {} with no log of the LLM's actual response. The ParsingError had a raw_output field that was never populated. Debugging was impossible.

  3. Prompt + LLM interaction: The magazine prompt was too permissive. Models (especially Qwen2.5) responded with Chinese preambles, markdown fences ( ```json), and tried to extract article content as nested JSON — hitting the 512-token limit and producing truncated/invalid JSON.