ai-citer 1.0.2

AI-powered fact extraction and citation mapping for documents (PDF, Word, web, text)

ai-citer

AI-powered fact extraction and citation mapping for documents — PDF, Word, web pages, and plain text.

Built on FastAPI + Anthropic Claude. Extracts verbatim-quoted facts from documents, maps each quote back to its exact character offset, and optionally assigns PDF page numbers.

Install

pip install ai-citer

Requires Python 3.11+ and a PostgreSQL database.

Quick start

Run as a standalone server

Set environment variables (or create a .env file):

ANTHROPIC_API_KEY=sk-ant-...
DATABASE_URL=postgresql://user:pass@localhost/ai_citer

ai-citer serve          # starts on :3001
ai-citer serve --port 8080 --reload

Or with uvicorn directly:

uvicorn ai_citer.main:app --port 3001

Embed the router in your own FastAPI app

from fastapi import FastAPI
from ai_citer import documents_router

app = FastAPI()
app.include_router(documents_router, prefix="/ai-citer")

Note: the router reads app.state.pool (asyncpg pool) and app.state.anthropic_client from the FastAPI app state. Use the lifespan from ai_citer.main as a reference, or set them up yourself.

Use the core functions directly

import anthropic
import asyncio
from ai_citer import (
    create_pool, init_db,
    extract_facts, map_citations, assign_page_numbers,
    parse_pdf, parse_word, parse_web, parse_text,
)

async def main():
    pool = await create_pool("postgresql://localhost/mydb")
    await init_db(pool)

    client = anthropic.AsyncAnthropic(api_key="sk-ant-...")

    # Parse a PDF
    with open("report.pdf", "rb") as f:
        content = parse_pdf(f.read())

    # Extract facts
    extraction, usage = await extract_facts(client, content.rawText)

    # Map quotes back to character offsets
    facts = map_citations(content.rawText, extraction.facts)
    print(facts[0].citations[0].charOffset)   # exact position in raw text
    print(f"Cost: ${usage.costUsd:.4f}")

asyncio.run(main())

REST API

When running as a server, the following endpoints are available under /api/documents:

Method	Path	Description
GET	/	List all documents
POST	/	Upload a file (multipart/form-data) or URL (url form field)
GET	/:id	Get a document (includes pdfData for PDFs)
POST	/:id/extract	Run fact extraction (optional { "prompt": "..." } body)
GET	/:id/facts	Get all accumulated facts for a document
POST	/:id/chat	Chat with a document ({ "message": "...", "history": [] })

MCP server

ai-citer ships an MCP server that exposes extraction tools to AI assistants (Claude Desktop, etc.):

ai-citer mcp

Tools: upload_document_url, extract_facts, get_facts, list_documents.

Environment variables

Variable	Required	Default	Description
ANTHROPIC_API_KEY	Yes	—	Anthropic API key
DATABASE_URL	Yes	—	PostgreSQL connection string

Development

git clone https://github.com/czawora/ai-citer
cd ai-citer/server
pip install -e ".[dev]"
pytest