docstore

dbt for unstructured data. Extract once, query forever.

Most tools re-read your documents every time you ask a question. docstore extracts the fields you care about once, caches them locally, and answers subsequent queries from the cache - no LLM calls, no re-reading, no waiting.

Run 1 (cold): 200 invoices → 400 LLM calls → ~$0.40
Run 2 (warm): 200 invoices → 0 LLM calls  → $0.00
Query:        "which invoices are unpaid?" → 0 LLM calls → <1s

---

Why this exists

LLMs made it easy to extract structured data from documents. What they did not provide is a layer that persists that extraction and invalidates it automatically when the source file changes. Every existing tool either:

• Re-reads raw documents on every query (expensive, slow)
• Requires a database or vector store (complex, overkill for most teams)
• Stores embeddings for semantic search (wrong abstraction for structured extraction)

docstore treats structured extraction as a cache over your unstructured data. Same insight as dbt applied to SQL - you define the transformation once, the system manages the state.

---

Benchmark

docstore ships with a reproducible public cache benchmark. It generates a synthetic invoice corpus, writes ground_truth.jsonl, then measures:

• cold_extract: empty cache, every document calls the LLM once
• warm_extract: same corpus and schema, every document is served from cache
• cached_query: filtered query against cached JSON — zero LLM calls, near-instant

# Synthetic invoices (auto-generated)
uv run python scripts/benchmark.py /tmp/docstore-benchmark --count 30
uv run python scripts/benchmark.py /tmp/docstore-benchmark --count 30 --filter "paid=false"

# Any existing folder of documents
uv run python scripts/benchmark.py ./my-documents \
  --no-generate --schema my_schema --ask --glob "*.pdf" --filter "status=active"

Use --provider and --model to run it against a specific vendor. The benchmark is intended to show cache behavior, not provider quality.

A contract-specific corpus generator is also included for demo purposes:

# Generate 30 synthetic vendor contract PDFs
uv run --extra scripts python scripts/generate_pdf_contracts.py ./sample_contracts

# Then benchmark against them
uv run python scripts/benchmark.py ./sample_contracts \
  --no-generate --schema contracts --ask --glob "contract_*.pdf" --filter "status=active"

---

Installation

pip install lumient-docstore

All four LLM providers (Anthropic, OpenAI, Groq, Gemini) work out of the box - pick one at runtime via --provider.

Or from source:

git clone https://github.com/LumientAI/docstore
cd docstore
pip install -e ".[dev]"

---

Quickstart

Python API

from pathlib import Path
from docstore import DocStore, ExtractionSchema, create_llm_client
from docstore.agents.orchestrator import run_directory

class InvoiceSchema(ExtractionSchema):
    vendor: str
    amount: float
    currency: str
    due_date: str
    paid: bool

invoices_dir = Path("./invoices")

# Co-locate the cache with the corpus so the CLI and Python API
# share state (the CLI's path-taking commands default to this).
store = DocStore(root=invoices_dir / ".docstore")
descriptor = InvoiceSchema.to_descriptor()
client = create_llm_client()  # defaults to Anthropic; pass provider="openai" etc. to override
results = run_directory(invoices_dir, descriptor, store, client)

# Query without any LLM calls
unpaid = store.query("InvoiceSchema", lambda r: r.data.get("paid") is False)

CLI

# Generate a synthetic invoice corpus for testing (30 .txt files)
python scripts/generate_txt_invoices.py ./sample_invoices --count 30

# Extract - describe fields interactively
docstore shell ./invoices/

# Extract with a named schema
docstore extract ./invoices/ --schema invoice_schema

# Use OpenAI, Groq, or Gemini instead of the default Anthropic provider
docstore extract ./invoices/ --schema invoice_schema --provider openai
docstore extract ./invoices/ --schema invoice_schema --provider groq
docstore extract ./invoices/ --schema invoice_schema --provider gemini

# Override the default model for any provider
docstore extract ./invoices/ --schema invoice_schema --provider gemini --model gemini-2.5-pro

# Query stored results (no LLM)
docstore query invoice_schema --filter "is_paid=false" --store ./invoices/.docstore

# Aggregate: count and sum per vendor (no LLM)
docstore query invoice_schema --group-by vendor --count --sum amount --store ./invoices/.docstore

# Ask in natural language - one LLM call compiles to a filter,
# results come from cache with no per-document re-reads
docstore ask "which unpaid invoices are over $5000?" --schema invoice_schema --store ./invoices/.docstore

# Diff a changed file
docstore diff ./invoices/acme_april.pdf --schema invoice_schema

# Remove cache entries whose source file no longer exists
docstore sync --store ./invoices/.docstore        # dry run
docstore sync --store ./invoices/.docstore --yes  # delete stale entries

# Wipe the cache (optional --schema X to scope)
docstore clean --store ./invoices/.docstore --yes

# Stats
docstore stats --store ./invoices/.docstore

MCP server (Claude Desktop)

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "docstore": {
      "command": "docstore-server",
      "env": {
        "DOCSTORE_DIR": "/path/to/your/.docstore",
        "DOCSTORE_PROVIDER": "anthropic",
        "ANTHROPIC_API_KEY": "your-key"
      }
    }
  }
}

Claude can then call extract, query, diff, and stats directly.

Supported providers are anthropic (default), openai, groq, and gemini. Set ANTHROPIC_API_KEY, OPENAI_API_KEY, GROQ_API_KEY, or GEMINI_API_KEY for the provider you choose. Each provider has a default model, and you can override it with --model on the CLI or DOCSTORE_MODEL for the MCP server.

---

How it works

┌──────────────────────────────────────────────────────────────────┐
│                         docstore pipeline                        │
│                                                                  │
│  document.pdf                                                    │
│       │                                                          │
│       ▼                                                          │
│  ┌─────────┐     cache hit?   ──────────────────────────────┐    │
│  │  Parser │  ─────────────►  .docstore/{key}.json          │    │
│  └─────────┘     (no LLM)     ──────────────────────────────┘    │
│       │                                                          │
│       │ cache miss                                               │
│       ▼                                                          │
│  ┌───────────┐                                                   │
│  │ Extractor │  1 LLM call - extract fields against schema       │
│  └───────────┘                                                   │
│       │                                                          │
│       │   (opt-in via --validate)                                │
│       ▼                                                          │
│  ┌╌╌╌╌╌╌╌╌╌╌╌┐                                                   │
│  ╎ Validator ╎  +1 LLM call - sanity-check extracted values      │
│  └╌╌╌╌╌╌╌╌╌╌╌┘                                                   │
│       │                                                          │
│       ▼                                                          │
│  .docstore/{file_hash}__{schema}__{version}.json                 │
└──────────────────────────────────────────────────────────────────┘

The validator is off by default - cold extraction is one LLM call per file. Pass --validate to add a plausibility check (doubles cost; see the CLI reference for trade-offs).

Cache key: sha256(file_bytes)[:16] + schema_name + sha256(json.dumps(fields, sort_keys=True))[:12]

The cache invalidates automatically when:

• The file content changes (file hash changes)
• The schema changes (schema version changes)
• A different schema is applied to the same file (different key)

---

Schema definition

Two ways to define a schema:

1. Python class (recommended for code)

from docstore import ExtractionSchema

class ContractSchema(ExtractionSchema):
    parties: list
    start_date: str
    end_date: str
    obligations: list
    auto_renews: bool

2. Natural language via CLI (recommended for ad-hoc use)

docstore shell ./contracts/
# > vendor name, contract start date, expiry date, whether it auto-renews

The orchestrator normalises your description into a canonical schema and shows it to you before running.

---

Supported file types

PDF, DOCX, TXT, MD, CSV, HTML, JSON

PDF support covers documents with embedded/selectable text. Scanned or image-only PDFs need OCR, which docstore does not support yet.

---

Limitations vs Lumient

docstore is a single-document extraction cache. It does not:

• Compose records across documents (invoice + Stripe → reconciliation status)
• Trigger automatically when files arrive
• Maintain a queryable entity layer with lineage
• Support multi-step workflow logic (validate, diff, generate)
• Provide governance and audit trails for regulated industries

For cross-document composition and maintained operational records, see Lumient.

---

Development

# uv (recommended)
uv sync --all-extras
uv run pytest
uv run ruff check .

# Or with pip
pip install -e ".[dev]"
pytest tests/

See AGENTS.md for architectural invariants and CONTRIBUTING.md for the PR process.

---

License

MIT