veco-ai 0.1.0

Vectorization & RAG Toolkit

veco-ai

veco-ai is a Python toolkit (Python 3.10-3.11) that converts a broad range of document types - text, PDF, Word, PowerPoint, images, audio, and video - into vector representations that can be queried through Retrieval Augmented Generation (RAG).
Embeddings are stored inside a FAISS index and can optionally be persisted to JSON (fallback), SQLite, or MongoDB. The integrated RAG interface lets you query knowledge bases via local Ollama models.

Features

• Automatic input detection for text, PDF, Word, PowerPoint, images, audio, and video
• Text extraction via pdfplumber, python-docx, python-pptx, pytesseract, moviepy, and whisper
• Speaker diarization (optional through veco_diarization.py)
• Vision extensions:
  • OCR via pytesseract
  • CNN classification (torchvision ResNet)
  • External image captioning (optional via veco_pic_describe)
• Chunking with overlap for RAG-ready embeddings
• Optional summaries generated with Ollama models (stored separately, never used as embedding input)
• FAISS index for efficient retrieval
• Persistence backends: JSON (fallback, stand-alone), SQLite, or MongoDB
• RAG queries: End-to-end helper (query()) that retrieves context and triggers an Ollama response

Project Structure

.
|-- veco_ai/
|   |-- __init__.py
|   |-- veco_ai.py              # Core vectorization library
|   |-- veco_diarization.py  # Optional speaker diarization pipeline
|   `-- veco_pic_describe.py # Optional image captioning helpers
|-- test/
|   `-- veco_test.py         # Example usage script
|-- requirements.txt
|-- pyproject.toml / setup.py
|-- test_data/               # Sample files for testing
|-- vector_db.json           # Example JSON database (fallback storage)
`-- UML/                     # Architecture diagrams

Dependencies

• Python 3.10 or 3.11
• torch, torchaudio, torchvision (CPU wheels via PyPI; follow the official PyTorch guide for CUDA)
• sentence-transformers
• faiss-cpu
• openai-whisper
• pdfplumber
• pytesseract
• pillow
• moviepy
• python-docx
• python-pptx
• numpy and scipy
• ollama
• webrtcvad-wheels, librosa, soundfile, speechbrain
• (See requirements.txt / pyproject.toml for exact versions)

Installation

1. Create a virtual environment

./setup_venv.ps1
# or
python3.11 -m venv .venv
.venv\Scripts\activate      # Windows
source .venv/bin/activate   # Linux/macOS

2. Install the base dependencies

pip install veco_ai

For local development instead of the published wheel, install from source:

pip install -r requirements.txt
# or
pip install -e .

3. Configure PyTorch (optional)

Follow the official PyTorch installation guide for your GPU/CPU setup.
For a CPU-only environment the default pip install from the requirements is sufficient.

Usage

Example script (tests/veco_test.py)

python tests/veco_test.py

The script loads or creates vector_db.json, vectorizes all files in the test_data/ folder, and saves the updated database.

Direct usage in Python

from veco_ai import Vectorize

# JSON fallback backend
veco = Vectorize(preload_json_path="vector_db.json")

# Vectorize a file
veco.vectorize("path/to/file.pdf", use_compression=True)

# Persist the database
veco.save_database("vector_db.json")

# Run a RAG query (Ollama required)
res = veco.query(
    database="vector_db.json",
    question="What is this document about?",
    llm_model="gemma3:12b",
)
print(res["answer"])

Architecture

The central class is Vectorize:

• Input detection: identifies the file type
• Text extraction: uses type-specific libraries
• Optional compression: generates summaries through Ollama
• Chunking: splits text into overlapping segments
• Embedding: performed with sentence-transformers
• Storage: FAISS index plus JSON/SQLite/MongoDB backends
• RAG: retrieves relevant context and optionally queries an Ollama model

Development

Install the development extras to run linting and tests:

pip install .[dev]
pytest

License

The project is released under the terms of CC0 1.0 Universal.