docqwise 0.3.1

Document intelligence that adapts, accelerates, and scales.

---

What is DocQWise?

DocQWise is a pluggable, AI-powered document intelligence engine. It reads any document format, extracts structured data using LLMs and RAG pipeline, and retrieves information with semantic search — locally, at scale, for zero per-page cost.

Install

Choose your install based on what you need:

# Option 1: Core only (PDF reading, regex extraction, no ML)
pip install docqwise

# Option 2: With ML (RAG pipeline, embeddings, OCR — recommended)
pip install -r requirements-ml.txt

# Option 3: Everything (all features, all formats)
pip install -r requirements-full.txt

LLM backend (pick one)

# Ollama — local, free, recommended
# Download from https://ollama.com then:
ollama pull nemotron-mini

# OR HuggingFace — local GPU
pip install transformers torch bitsandbytes accelerate

# OR OpenAI — cloud API
export OPENAI_API_KEY=your-key

Quick Start

from docqwise import Docqwise

dq = Docqwise()

# Ingest any document
dq.ingest("documents/")

# Extract fields with template
result = dq.extract_fields("invoice.pdf", template="invoice")
print(result.to_json())

# Ask questions about structured data
dq.ingest("sales.csv")
answer = dq.ask("What is the total amount?")

Extraction Methods

dq = Docqwise()

# RAG (default) — chunk → embed → retrieve → LLM extract
dq.extract_fields("doc.pdf", template="invoice")

# Direct LLM
dq.extract_fields("doc.pdf", template="invoice", method="llm")

# Vision (scanned docs, handwriting)
dq.extract_fields("scan.jpg", method="vision", model="gpt-4o")

# Regex (fast, no ML)
dq.extract_fields("doc.pdf", template="invoice", method="regex")

LLM Backends

# Ollama (local)
dq.extract_fields("doc.pdf", model="nemotron-mini")

# HuggingFace (local GPU)
from docqwise.llm.hf_llm import HuggingFaceLLM
llm = HuggingFaceLLM("Qwen/Qwen2.5-3B-Instruct", quantize="4bit")
dq.extract_fields("doc.pdf", llm=llm)

# OpenAI (cloud)
dq.extract_fields("doc.pdf", model="gpt-4o-mini")

Templates

dq.extract_fields("invoice.pdf", template="invoice")
dq.extract_fields("contract.pdf", template="contract")
dq.extract_fields("resume.pdf", template="resume")
dq.extract_fields("receipt.jpg", template="receipt")

# Custom schema
schema = {
    "vendor": {"type": "string", "description": "Company name"},
    "total": {"type": "number", "description": "Total amount"},
}
dq.extract_fields("doc.pdf", schema=schema)

Custom Prompts

You design the prompts. We run the pipeline.

dq = Docqwise()

# Default — docqwise handles the prompt
dq.extract_fields("doc.pdf", template="invoice")

# Your own prompt — full control
dq.extract_fields("doc.pdf", prompt="""
You are a medical record parser.
Extract patient name, diagnosis, and prescribed medications.
Return JSON only.

Document:
{context}

JSON:
""")

# Your prompt template with schema
dq.extract_fields("doc.pdf",
    schema={"patient": {"type": "string"}, "diagnosis": {"type": "string"}},
    prompt_template="""
Given this extraction schema:
{schema}

Parse this document:
{context}

Return JSON matching the schema exactly.
""")

System Prompt

Set the LLM's role before extraction. Works with all backends and all RAG modes.

# Set the role — LLM behaves as this specialist
dq.extract_fields("invoice.pdf",
    system_prompt="You are a maritime document specialist with 20 years experience.",
    prompt="Extract vessel name, port, and total cost.

Document:
{context}

JSON:",
)

# System prompt + schema + template — full control
dq.extract_fields("contract.pdf",
    system_prompt="You are a legal analyst specializing in shipping law.",
    schema={"parties": {"type": "array"}, "liability": {"type": "number"}},
    prompt_template="Extract {schema} from:
{context}
JSON:",
)

# System prompt with GraphRAG
dq.ask_rag("What is the liability cap?",
    mode="graphrag",
    system_prompt="You are a risk analyst. Be precise with numbers.",
)

Self-Improving Corrections

result = dq.extract_fields("invoice.pdf", template="invoice")
result.correct({"tax": 33300.00, "gst_number": "29AABCU9603R1ZM"})
# Next similar document → corrections applied automatically

Structured Data Q&A

dq.ingest("sales.xlsx")
dq.ask("What is the total amount?")         # exact SUM
dq.ask("Which vendor has highest sales?")    # GROUP BY + MAX
dq.ask("How many invoices are overdue?")     # COUNT + WHERE

Source Attribution

Every answer tells you which document it came from.

# Simple ask — shows source in output
answer = dq.ask("What is the total amount?")
print(answer)
# 2,18,300.00
#   Source: sample_invoice.pdf

# Detailed answer with metadata
result = dq.ask_with_source("What is the governing law?")
print(result.answer)        # "laws of India"
print(result.source_name)   # "sample_contract.pdf"
print(result.confidence)    # 0.85
print(result.method)        # "llm"

RAG Modes

Choose your retrieval strategy:

dq = Docqwise()
dq.ingest("documents/")

# General RAG (default) — chunk → embed → retrieve → LLM
result = dq.ask_rag("total amount?", mode="general")

# GraphRAG — entity graph → graph traversal → cross-document answers
result = dq.ask_rag("Who is the vendor for PO-2012?", mode="graphrag")
print(result["answer"])         # "SuperStore"
print(result["source"])         # "invoice_001.pdf"
print(result["evidence"])       # [{entity, relation, target}, ...]

# Multimodal RAG — text + images → vision LLM
result = dq.ask_rag("What is in this scan?", mode="multimodal", source="scan.jpg")

Graph Visualization

Build and visualize document knowledge graphs:

dq = Docqwise()
dq.ingest("documents/")

# Build knowledge graph from all documents
graphrag = dq.build_document_graph()
graph = graphrag.get_graph()
print(f"Nodes: {graph.node_count}, Edges: {graph.edge_count}")

# Generate interactive HTML visualization
dq.visualize_graph(output="my_graph.html")
# Open my_graph.html in browser — interactive, color-coded, draggable nodes

All Features

dq = Docqwise()

# Ingestion
dq.ingest("file.pdf")                    # single file
dq.ingest("documents/")                  # folder (all formats)
dq.ingest("data.csv")                    # structured data

# Extraction
dq.extract_fields("doc.pdf")             # field extraction
dq.extract_tables("doc.pdf")             # table extraction
dq.extract_entities("doc.pdf")           # entity extraction
dq.extract_images("doc.pdf")             # image extraction
dq.extract_text("doc.pdf")               # text extraction
dq.auto_extract("doc.pdf")               # auto-detect + extract

# Intelligence
dq.retrieve("query", top_k=5)            # semantic search
dq.ask("question")                       # Q&A
dq.classify("doc.pdf")                   # classification
dq.compare("v1.pdf", "v2.pdf")           # comparison
dq.detect_schema("data.csv")             # schema detection
dq.detect_pii("doc.pdf")                 # PII detection

Demos

Run in order:

Demo	What	Install
python demo/01_quickstart.py	All core features	pip install docqwise
python demo/02_ollama.py	AI extraction with Ollama	ollama pull nemotron-mini
python demo/03_huggingface.py	AI extraction on GPU	pip install transformers torch bitsandbytes accelerate
python demo/04_rag.py	Full RAG pipeline	pip install sentence-transformers

Notebook

pip install jupyter
jupyter notebook notebooks/docqwise_getting_started.ipynb

Testing

pip install pytest
pytest -v

Docker

docker compose up --build

Architecture

engine.py (stable — never changes)
    └── factory.py (all component selection)
            ├── ExtractorFactory  → rag | llm | vision | regex
            ├── LLMFactory        → ollama | huggingface | openai
            ├── EmbedderFactory   → sentence-transformers | any
            ├── StoreFactory      → sqlite | qdrant | faiss | any
            ├── ChunkerFactory    → structure | fixed | sentence
            └── TemplateFactory   → invoice | contract | resume | receipt

Ecosystem

Library	Tagline	Domain
SightRAG	See. Search. Retrieve.	Visual intelligence
sonarwise	Hear. Search. Retrieve.	Audio intelligence
docqwise	Read. Extract. Retrieve.	Document intelligence
adaptive-intelligence	Learn. Remember. Adapt.	Orchestration
llmevalkit	Evaluate. Score. Improve.	Evaluation

License

Apache License 2.0

Author

Venkatkumar Rajan