Structured Data Extraction

Pull user-specified fields out of arbitrary PDFs and return them as structured JSON with provenance.

It's domain-agnostic: nothing domain-specific is checked in. The vocabulary, acronyms, and field shapes are all derived from each document at runtime, so the same pipeline handles a chip datasheet or an earnings report without changes.

How it works

Give it a PDF and a list of keys. It converts the PDF to text, learns the document's vocabulary, splits it into typed sections, and embeds them. Then it finds the sections and rows closest to your keys, cleans up any messy tables among them, and asks an LLM to read just those passages and return structured values.

PDF
 └─► convert  (docling)            → markdown + items.json
  └─► vocab  (YAKE)                → vocab.json
   └─► classify (RandomForest)     → sections.json
    └─► tables (complexity score)  → tables.json
     └─► embed (sentence-transformers) → *_embeddings.npy + row_map
      └─► search (cosine similarity)   → matches.json
       └─► refine matched tables (vision LLM, lazy) → cleaned HTML
        └─► extract (schema + LLM) → output/<stem>_extraction_<model>.json

Each step writes its output to disk and is skipped on a re-run if the file already exists, so iterating on a prompt does not re-run docling.

Why this design

• Vocabulary is learned per document. vocab.py runs YAKE keyword extraction on each PDF instead of using a fixed word list, and a one-word key like gain is expanded with the document's related terms (gain error, gain drift) before search.
• Search is semantic. "CMRR" and "common-mode rejection ratio" share no letters but sit close in embedding space, so cosine similarity over sentence-transformer embeddings finds the passage where string matching would miss it.
• Table rows are embedded one at a time, each with its group header prepended, so a row reads like Gain Error > G=1 | 0.01 | 0.04 % instead of a bare 0.01 | 0.04 %.
• Vision runs lazily. The vision LLM only cleans up a table if that table matched a key, so cost tracks what you asked for rather than document size.
• The extraction LLM is constrained. extract.py discovers a per-key schema, builds a Pydantic model from it, and hands it to Instructor, which validates the output instead of accepting free text.
• Running several models is optional. Extraction fans out across the configured LLMs in parallel; records are de-duplicated (table-sourced ones preferred) and rater.py scores their agreement and attaches a confidence to each record. With one model, every record is confidence 1.0.

Pipeline at a glance

Supporting: matching.py (same-fact equivalence), features.py (classifier feature vector), paths.py (cache filename conventions), config.py (loads .env, derives paths, auto-selects CUDA/MPS/CPU).

Design notes

Each Python file has a module-level docstring explaining what it does and the non-obvious design decisions behind it. Start with the docstring at the top of the file you want to understand.

Installation

Requirements: Python 3.11+, plus an OpenAI-compatible LLM endpoint (the defaults assume a local Ollama). The first run also downloads the embedding model from Hugging Face.

# 1. create and activate a virtual environment
python3 -m venv venv
source venv/bin/activate

# 2. install from PyPI
pip install structured-data-extraction

To modify the pipeline itself, install an editable checkout instead:

git clone <repo-url> && cd structured-data-extraction
pip install -e .

Create a .env in the directory you run the tool from. It is required: the package reads these values on startup and will not run without them. It searches the working directory and its parents, and real shell environment variables override the file. Copy this block and edit the values:

# Provider: any OpenAI-compatible endpoint (Ollama, OpenAI, ...)
LLM_BASE_URL=http://localhost:11434/v1
LLM_API_KEY=ollama

# Models
SCHEMA_MODEL=glm-5.1:cloud
EXTRACTION_MODELS=gemma4:31b-cloud      # comma-separated for parallel voting
VISION_MODEL=kimi-k2.7-code:cloud       # must have vision abilities
EMBEDDING_MODEL=nomic-ai/nomic-embed-text-v1.5  # HugginFace model ID (downloaded locally on the first run)

# Tuning
EMBED_BATCH_SIZE=4        # small so long tables don't OOM a local GPU; raise on bigger hardware
MAX_LLM_RETRIES=2         # per-model retry budget on extraction failures
EXTRACTION_WORKERS=4      # parallel workers when running multiple extraction models

# Pipeline tuning = OPTIONAL: omit this whole block and the defaults below apply.
VOCAB_TOP_N=50                    # YAKE vocab terms kept per document
VOCAB_MAX_NGRAM=4                 # longest phrase YAKE treats as one term
SIMILARITY_THRESHOLD=0.55         # cosine floor for a section/row to count as a match
SEARCH_TOP_K=25                   # max matches kept per key (above the threshold)
VISION_COMPLEXITY_THRESHOLD=0.15  # min table complexity worth sending to the vision model
VISION_PARALLELISM=4              # parallel vision calls when refining matched tables
EXPAND_KEYS=true                  # expand each key with the document's own related vocab before search

Notes:

• EXTRACTION_MODELS is a comma-separated list. List several to cross-check agreement, or one to skip the agreement step. The first model is the "primary", used wherever a single model is needed (such as text-based table cleanup).
• For a non-Ollama provider, set LLM_BASE_URL / LLM_API_KEY accordingly (for OpenAI: https://api.openai.com/v1 and sk-...).
• Shell environment variables override the file. Make sure every named model actually exists on your provider, or that step retries and then fails.
• The pipeline tuning block is optional: every value falls back to the default shown above if it's absent from .env. They are the knobs for recall vs. cost — e.g. lower SIMILARITY_THRESHOLD or raise SEARCH_TOP_K to retrieve more (and feed the LLM more), raise VISION_COMPLEXITY_THRESHOLD to send fewer tables to the (slow) vision model.

Usage

CLI:
# one PDF
structured-extract pdf/1167fc.pdf --keys gain CMRR "supply voltage" #when key is more than one word then it should be inside ""

# a whole directory
structured-extract pdf/ --keys revenue "net income"

Script:
from pathlib import Path
from structured_data_extraction.main import process_single_pdf, process_batch, finalize_outputs

path = Path("pdf/example.pdf")
# OR dir = Path("pdf_dir")
keys = ["key_1", "key_2", "key_3"]

results = process_single_pdf(path, keys)
# OR process_batch(pdf_dir, user_keys)



finalize_outputs(levels=["all"])

print("Done. See output/fused.json and output/records.json")

Wrap multi-word keys in quotes. If you omit --keys, it prompts for a comma-separated list.

Custom schema (skip schema discovery)

Pass --schema with a JSON file to bypass the schema-discovery LLM call entirely. The file must contain document_type and schema keys, using the same format as intermediate/*_schema.json. Keys are derived from the schema, so --keys is optional.

Use "fixed" on any sub-field to constrain the LLM to a single value — the field becomes a Literal type in the Pydantic model, so the model can only emit that exact value. This lets you slice a table down to specific variants, grades, or qualifiers:

structured-extract pdf/AD840.pdf --schema my_schema.json

{
  "document_type": "operational amplifier datasheet",
  "schema": {
    "quiescent_current_s_typ": {
      "type": "list",
      "description": "Quiescent current typical value for the AD840S variant",
      "fields": {
        "differentiator": {
          "type": "string",
          "description": "Product variant",
          "fixed": "AD840S"
        },
        "qualifier": {
          "type": "string",
          "description": "Column label: min, typ, or max",
          "fixed": "Typ"
        },
        "condition": { "type": "string", "description": "Operating condition" },
        "value": { "type": "number", "description": "Quiescent current value", "unit": "mA" },
        "metadata": { "type": "object", "description": "Source and provenance" }
      }
    }
  }
}

A full run writes per-model extractions to output/<stem>_extraction_<model>.json, then fuses them into fused.json (every fact, sorted by confidence) plus rating_report.json, and wraps fused.json into the provenance-stamped records.json.

A multi-model run then ends by printing the confidence-tier histogram and offering to keep a subset of tiers — pass --levels (e.g. --levels 1.0 0.75, or all) to choose up front, or omit it for an interactive prompt. Single-model runs skip this step, since every record is confidence 1.0.

Re-run a single stage. Every step is also available as its own console script. Run from the same working directory as the full pipeline:

# early stages (operate on files from the previous step)
structured-extract-convert --input pdf/example.pdf        # PDF to markdown
structured-extract-vocab --input markdown/example.md      # domain vocabulary
structured-extract-classify --stem example                # section classification
structured-extract-embed --stem example --keys gain CMRR  # embed sections + keys

# retrieval and extraction
structured-extract-search --stem example --original-keys gain CMRR  # semantic search
structured-extract-extract --stem example --keys gain CMRR          # LLM extraction

# post-processing (no --stem flag, they read output dir)
structured-extract-rate                          # fuse models, score agreement
structured-extract-records                       # wrap fused.json into records.json
structured-extract-filter --levels 1.0 0.75     # keep only selected confidence tiers

Confidence is models-agreeing / total-models. filter_records reads records.json and keeps any subset of the confidence tiers present in the data — pass --levels (e.g. 1.0 0.75, or all), or omit it for an interactive menu. Needs more than one model to be meaningful.

Outputs

Every stage caches its result under intermediate/ and is skipped on a re-run if the file already exists; the finished deliverables land in output/.

intermediate/<stem>_items.json          docling items (label, page, bbox)
              <stem>_vocab.json          YAKE vocabulary
              <stem>_sections.json       classified content stream
              <stem>_tables.json         tables + complexity scores
              <stem>_*_embeddings.npy    section / key vectors
              <stem>_matches.json        search hits per key
              <stem>_run.json            run manifest (models, timings)

output/<stem>_extraction_<model>.json    one file per extraction model
       fused.json                        every fact, confidence-sorted per key  ← the deliverable
       records.json                      fused.json wrapped in a provenance envelope, one per document
       rating_report.json                agreement scoring + pairwise Cohen's kappa

A records.json entry pairs a provenance header with the extracted fields. Each field is a list of records sorted by confidence (models-agreeing / total-models), so the top record per key is the one the most models agreed on:

{
  "0150200051": {
    "type": "ExtractionResult",
    "provenance": {
      "source_file": "0150200051.pdf",
      "document_type": "cable connector product datasheet",
      "models": { "schema_discovery": "glm-5.1:cloud", "vision": "kimi-k2.7-code:cloud",
                  "embedding": "nomic-ai/nomic-embed-text-v1.5" },
      "user_keys": ["product family", "cable length", "Voltage", "net weight"]
    },
    "product_family": [
      { "confidence": 1.0,  "models": ["deepseek-v4-flash:cloud", "gemma4:31b-cloud",
                                       "gpt-oss:120b-cloud", "minimax-m2.5:cloud"],
        "value": "Cable",        "metadata": { "section_type": "text", "specified": true } },
      { "confidence": 0.25, "models": ["deepseek-v4-flash:cloud"],
        "value": "Cable Series", "metadata": { "section_type": "text", "specified": true } }
    ],
    "voltage": [
      { "confidence": 1.0, "models": ["deepseek-v4-flash:cloud", "gemma4:31b-cloud",
                                      "gpt-oss:120b-cloud", "minimax-m2.5:cloud"],
        "qualifier": "Maximum", "value": 60.0, "value_unit": "V AC",
        "metadata": { "section_type": "text", "specified": true } }
    ]
  }
}

Testing

pip install -e ".[dev]"
pytest

License

MIT, see LICENSE.