mineru-refine 0.12.0

MinerU output linter/fixer — LLM tool-use loop that restructures (never generates) MinerU content_list. Machine-verified fidelity: C_out ⊆ C_in.

mineru-refine

🌏 中文文档：README.zh-CN.md

A post-processor (linter / fixer) for MinerU parse results.

It takes MinerU's content_list (an array of item objects), fixes the high-frequency structural problems that parsing introduces — pseudo-headings, cross-page sentence breaks, cross-page split tables, headers/footers mixed into the body, LaTeX / link residue — and returns a content_list with the same schema, so downstream consumers need zero changes.

Two core promises:

• Never adds a single character: it only removes and reorganizes; every content character in the output comes from the input, verified step by step by machine, and any violation is rolled back automatically (not enforced by prompting the LLM).
• fail-open: any exception / LLM unavailability → the input is returned unchanged (report["failOpen"] == True), never breaking the upstream pipeline.

This package is a native PyO3 binding of the Rust core implementation; its options and return values are structurally identical to the JS / Rust / HTTP versions.

Install

pip install mineru-refine

Requires Python ≥ 3.9.

Usage

import json
import mineru_refine

items = json.load(open("content_list.json"))

result = mineru_refine.refine(
    items,                              # content_list (list[dict])
    sha256="...",                       # optional: source-file SHA256; enables the in-process cache
    max_iterations=None,                # optional: hard cap on the fix loop; defaults to adaptive with suspect count
    concurrency=8,                      # optional: number of suspects judged in parallel; 1 = strictly serial
    image_dir="/abs/mineru/out",        # optional: MinerU output dir; enables vision judging for split tables
    fix_ocr_confusion=False,            # optional: opt-in OCR character-confusion fix layer (CE0→CEO, etc.)
    extra_confusion_pairs=None,         # optional: extra allowlist pairs for confusion, e.g. ["0D"]
    rewrite_garbled_tables=False,       # optional: opt-in vision re-transcription layer for garbled tables (needs image_dir)
    degrade_garbled_tables=False,       # optional: opt-in fallback that demotes unrecoverable garbled tables to images
    model_config=None,                  # optional: config-driven model swap (see "Swapping models" below)
    chat=None,                          # optional: custom text-LLM callback (escape hatch; takes priority over model_config)
    vision=None,                        # optional: custom vision-LLM object (escape hatch)
)

result["items"]    # cleaned content_list (same schema; unknown fields passed through verbatim)
result["report"]   # audit report: iterations / opCounts / dismissed / removedSpans
                   #               / violations / tokenUsage / failOpen
                   #               (with fix_ocr_confusion on, also confusionFixes etc.; see the main README)

Every removed span is recorded in report["removedSpans"] (itemId / original text / reason), auditable line by line. fix_ocr_confusion=True turns on the confusion-fix layer (direct replacement, LLM proposal + mechanical gates); once on, the output contract changes from "remove-only" to a dual contract — see the "Confusion-fix layer" section of the main README. rewrite_garbled_tables=True turns on vision re-transcription for heavily garbled tables (mechanical detection selects tables judged whole-table junk, Qwen-VL re-transcribes cell by cell against the crop, all recorded in report["tableRewrites"]) — see the "Garbled-table re-transcription layer" section of the main README. degrade_garbled_tables=True turns on the garbled-table fallback (purely mechanical, runs after the re-transcription layer: tables still judged junk that have an img_path are demoted to image, counted in report["tableDegraded"]) — see the "Fallback demotion" section of the main README.

Standalone helper functions (none call the LLM):

mineru_refine.render_markdown(items)    # items → full.md text (deterministic re-render)
mineru_refine.detect_suspects(items)    # detect suspects only, returns the suspect list

Swapping models (custom LLMs)

By default the text role is DeepSeek and the vision role is Qwen-VL. Two mechanisms let you point them at any other LLM (e.g. MiniMax); see the main README for the full explanation.

1. model_config — config-driven, multi-vendor (recommended). A dict with two independent roles (reasoning for text, vision for vision); each is {provider?, model, key?, baseUrl?} (camelCase keys). Omit a role to fall back to the env default.

mineru_refine.refine(
    items,
    image_dir="/abs/mineru/out",
    model_config={
        # MiniMax-M3 is OpenAI-compatible and multimodal, so one model serves both roles
        "reasoning": {"provider": "openai", "model": "MiniMax-M3", "key": key, "baseUrl": "https://api.minimaxi.com/v1"},
        "vision":    {"provider": "openai", "model": "MiniMax-M3", "key": key, "baseUrl": "https://api.minimaxi.com/v1"},
    },
)

provider accepts deepseek / aliyun (qwen, dashscope) / openai (openai-compatible, custom) / anthropic (claude) / gemini (google) / ollama / groq / xai (grok); omit it to infer from the model name.

2. Custom callbacks — the escape hatch (take priority over model_config). Pass your own implementations when you need auth/proxy/model logic model_config can't express:

def chat(messages, tools):
    # messages: list[dict] (OpenAI-style), tools: list[dict]
    return {
        "message": {"content": "...", "tool_calls": [...]},  # tool_calls optional
        "finishReason": "stop",
        "usage": {"prompt_tokens": 0, "completion_tokens": 0},
    }

class Vision:
    def judge_split_table(self, img_a: bytes, img_b: bytes):
        return {"verdict": "merge", "reason": "...", "usage": {}}   # verdict: "merge" | "dismiss"
    def transcribe_table(self, img: bytes, cells_render: str):      # optional; only for rewrite_garbled_tables
        return {"cells": [{"row": 0, "col": 0, "text": "..."}], "usage": {}}

mineru_refine.refine(items, image_dir="/abs/mineru/out", chat=chat, vision=Vision())

Environment variables

Variable	Required	Purpose
DEEPSEEK_APIKEY	Yes	Text judging (DeepSeek). If missing, refine goes straight to fail-open
QWEN_APIKEY	For vision judging	Qwen-VL judging of cross-page split tables; if missing, those suspects are skipped and the tables kept as-is

The library itself does not read .env; set the environment variables in the host program (or load .env yourself).

Building locally

just py-dev        # repo root: build the wheel and install it into bindings/python/.venv
just publish-py    # publish to PyPI: current-platform wheel + sdist (needs MATURIN_PYPI_TOKEN)

The full design docs for the detector, the fix operation set, and the fidelity gates are in the repository README.

License

MIT