Doco — Local Document Intelligence Platform

Doco is a privacy-first, fully local document intelligence platform inspired by LandingAI. It enables high-fidelity ingestion, structured data extraction, and semantic querying of PDFs and images—all without calling any external APIs.

By combining layout-aware VLM OCR pipelines with self-correcting agentic JSON extraction and hybrid RAG, Doco runs entirely on local hardware, keeping sensitive documents 100% secure.

---

Interface Preview

Here is the Doco workspace in action:

---

Key Features

• High-Fidelity Document Processing: Ingests multi-page PDFs and images using SuryaOCR (running via llama.cpp) for layout analysis, bounding box coordinates, reading-order alignment, and high-accuracy text recognition.
• Agentic JSON Extraction:
  • Interactive Schema Builder: Manually edit, upload a custom .json schema, or query the local VLM to automatically suggest a schema based on the document's structure.
  • Self-Correcting Critique Loop: Validates LLM extractions against the target JSON schema using jsonschema. If validation fails, it feeds the exact parser errors back to the model for correction (up to 3 attempts).
  • Threshold-Based Routing: Automatically routes documents based on character count to optimize processing paths (Direct VLM Extraction vs Chunked fallbacks).
• Local RAG Chat Interface:
  • Hybrid Search: Leverages a combined vector search (FAISS) and keyword retrieval (BM25) ensemble retriever.
  • Cross-Encoder Re-ranking: Uses ms-marco-MiniLM-L-6-v2 to re-rank chunks for high-relevance search context.
  • SSE Streaming: Answers user questions using Server-Sent Events (SSE) for real-time token-by-token streaming in the UI.
• Dynamic Memory Optimization: Automatically loads and unloads heavy Surya OCR models and Ollama services to prevent memory leaks and run efficiently on standard consumer hardware.

---

Technology Stack

• Backend: Python 3.11+, FastAPI, LangChain, Pydantic, jsonschema, PyPDFium2, FAISS, rank-bm25, SentenceTransformers.
• Local Models:
  • SuryaOCR (OCR, layout detection)
  • Ollama (qwen2.5vl:7b, qwen3-embedding:0.6b, glm-ocr)
• Frontend: Vanilla HTML5, CSS3, JavaScript (SSE streaming, JSON validator, responsive panes).

---

Getting Started

1. Prerequisites

Ensure you have Ollama installed on your system. Pull the required models:

ollama pull qwen3-embedding:0.6b
ollama pull qwen2.5vl:7b

2. Ingest dependencies

Clone the repository and set up a Python virtual environment:

# Set up virtual environment
python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

# Install packages
pip install -r requirements.txt

3. Configuration (Optional)

The platform uses sensible defaults, but you can override models and other settings using an environment file. Create a .env file in the root directory:

# Example .env overrides
LLM_MODEL=qwen2.5vl:7b
EMBEDDING_MODEL=qwen3-embedding:0.6b
RERANKER_MODEL=cross-encoder/ms-marco-MiniLM-L-6-v2
EXTRACTION_THRESHOLD=30000

4. Start the Platform

Run the FastAPI development server:

python main.py

Open your browser and navigate to http://localhost:8000/.

---

Roadmap

• Map-Reduce Chunked Extraction: Fully implement map-reduce aggregation for extracting schemas from massive documents that exceed VLM context boundaries.
• Multi-Document Indexes: Run cross-document comparisons and search queries across the entire processed document library.