Integration quickstart

VerifyDoc adds a per-field trust layer (calibrated confidence + source grounding + accept/review) on top of whatever extractor or framework you already use. None of these frameworks is a hard dependency — the integrations work by duck typing.

Instructor / Outlines / Marvin / Pydantic-AI (any BaseModel)

Verify an object that any structured-extraction framework produced:

from verifydoc.integrations.instructor import verify_instructor_result

obj = my_framework_extract(document_text)      # -> a pydantic.BaseModel instance
result = verify_instructor_result(document_text, obj, threshold=0.8)

for f in result.fields:
    print(f.path, f.value, round(f.confidence, 2), f.decision)   # accept / review

Pydantic-AI (or plain Pydantic)

from pydantic import BaseModel
from verifydoc import verify_model

class Invoice(BaseModel):
    invoice_id: str
    total: float

result = verify_model("invoice.pdf", Invoice, threshold=0.8)
print(result.to_dict())

LangChain

from verifydoc.integrations.langchain import VerifiedExtractor

extractor = VerifiedExtractor(chain.invoke, schema=Invoice, threshold=0.8)
result = extractor(document_text)              # -> VerifiedResult

LlamaIndex / DSPy / Haystack

Wrap your str -> dict | BaseModel extraction step and hand the output to verify_instructor_result (BaseModel) or verify(text, schema=...) (dict). See runnable scripts in examples/: - examples/pydantic_ai_example.py - examples/llamaindex_example.py

Agents / IDEs (MCP)

Claude Code, Claude Desktop, Cursor, Cline, Codex — point them at the verifydoc-mcp stdio server. Copy-paste configs: examples/mcp/README.md.

Self-hosted API / web / messaging

Run the FastAPI server for a REST /verify endpoint, a review web app, and WhatsApp/Telegram bots — all on your own infra. See docs/DEPLOY.md.