citefield

Extract metadata from technical manuals, with a citation for every value.

citefield is a Python library and CLI that reads a technical manual, extracts the fields you define in a YAML config, and returns JSON in which every value cites a verbatim quote and its page. A human can then review the result value by value.

Status: early development. What works today: PDF manuals; string, enum and number fields, with quotes verified and units converted in code; evaluation against hand-reviewed outputs. Planned: tables as evidence, Markdown input, field proposals and re-extraction that keeps earlier reviews. The plan is in the design.

Why

LLMs are good at finding values in long documents and bad at proving where those values came from. citefield keeps the LLM's job narrow:

1. The LLM only proposes. Code verifies, humans decide.
2. No evidence, no value. Every value cites a quote and its page, and code checks that the quote really appears in the document.
3. Deterministic where possible. Unit conversion, evidence checks and merging are code, not prompts.
4. Reviews are kept. Human decisions are stored in the output and become the evaluation set.

Quick start

Install the CLI (citefield is on PyPI from the first release, 0.1.0; until then, install from source):

uv tool install citefield   # or: pipx install citefield

citefield calls the model through LiteLLM, which reads the provider's API key from the environment. The example configs use Claude Haiku 5.5 on OpenRouter:

export OPENROUTER_API_KEY=sk-or-...

Define the fields you want:

fields:
  - name: dry_weight
    description: Dry weight of the generator set (less basic issue items), one per model.
    type: number
    unit: kg

settings:
  llm:
    model: openrouter/anthropic/claude-haiku-5.5
    temperature: null # not accepted by the model
  extraction:
    number_locale: en # the manual writes 1,140 for 1140

Run the extraction:

citefield extract tm-9-6115-642-10.pdf -c fields.yaml -o out/

Get values with their evidence and confidence signals:

{
  "name": "dry_weight",
  "status": "found",
  "values": [
    {
      "status": "derived",
      "raw_value": "1,140",
      "raw_unit": "lb.",
      "value": 517.1,
      "unit": "kg",
      "derivation": "1,140 lb. = 517.1 kg",
      "context": "MEP-803A",
      "evidence": [
        { "kind": "quote", "page": 27, "quote": "Dry Weights (less Basic Issue Items):" },
        { "kind": "quote", "page": 27, "quote": "MEP-803A 1,140 lb. (516.4 kg)" }
      ],
      "signals": {
        "evidence_match": 1.0,
        "citation_count": 1,
        "run_agreement": null,
        "model_confidence": 0.9,
        "unit_resolved": true
      }
    }
  ]
}

The getting-started guide walks through the whole loop on a manual from the test corpus: checking the config and the parsed text, extracting, reading the output, reviewing it and measuring accuracy. examples/ has configs for every corpus manual.

License

Apache 2.0. The manuals in corpus/ keep their own licenses, listed there. The design is inspired by Luca Bianchi's workshop Documents Are Not Knowledge (Codemotion Milan 2026); see the acknowledgements.