SightRAG

See. Search. Retrieve.

Quick Start

from sightrag import SightRAG

rag = SightRAG()
rag.index("./photos/")
results = rag.query("find empty shelf")
rag.show(results)

Install

pip install sightrag

For faster inference:

pip install sightrag[onnx]               # 2x faster (any CPU)

For v0.4 features:

pip install sightrag[ocr]                # reads text on images
pip install sightrag[multimodal]         # LLM understanding
pip install sightrag[grounding-dino]     # any domain detection
pip install sightrag[reid]               # person tracking
pip install sightrag[cli]                # terminal commands
pip install sightrag[qdrant]             # large scale store
pip install sightrag[all]                # everything

What's New in v0.4

• OCR integration : reads text on products, signs, labels, documents during indexing. "find Calgon 2kg" matches actual text on packaging. Zero query-time overhead.
• Multimodal LLM : optional semantic understanding at query time. Local models (Qwen2-VL) or API (GPT-4o). Use understand=True for deep queries.
• Text + visual hybrid search : OCR text match boosts CLIP visual scores automatically.
• All v0.3 features : Grounding DINO, Person Re-ID, CLI, Qdrant, re-ranking.
• Backward compatible : v0.1/v0.2/v0.3 code works unchanged.

What SightRAG Is

SightRAG is not a model. Not a wrapper. Not a framework plugin.

It is a complete visual retrieval system. You provide any detection model, any embedding model, any vector store. SightRAG handles the pipeline: load, detect, embed, index, retrieve.

All models and indexes are stored in ~/.sightrag/ : your project folder stays clean.

Project Structure

sightrag/
├── sightrag/
│   ├── core.py                  ← SightRAG main class
│   ├── backends/                ← auto-select fastest inference
│   │   ├── pytorch_backend.py   ← default (works everywhere)
│   │   ├── onnx_backend.py      ← 2x faster (any CPU)
│   │   ├── tensorrt_backend.py  ← fastest (NVIDIA GPU)
│   │   └── openvino_backend.py  ← Intel CPU optimized
│   ├── detectors/base.py        ← plug custom detection model
│   ├── embedders/base.py        ← plug custom embedding model
│   ├── store/                   ← SQLite (default) + ChromaDB
│   ├── visualizer.py            ← rag.show() with bounding boxes
│   ├── indexer.py               ← C++ core with Python fallback
│   ├── retriever.py             ← text + reference queries
│   └── api.py                   ← REST API (FastAPI)
│
├── cpp/                         ← C++ speed core (optional)
├── demo_sightrag/               ← demo scripts + test data
│   ├── sightrag_images.py       ← image folder demo
│   ├── sightrag_video.py        ← video indexing demo
│   ├── sightrag_livecam.py      ← webcam demo
│   ├── sightrag_restapi.py      ← REST API demo
│   ├── input_images/            ← sample images
│   └── reference_images/        ← reference query images
├── notebooks/                   ← Colab notebook
├── tests/                       ← unit tests
└── docs/                        ← documentation

How To Test

# Quick test
python test_sightrag.py

# Demo scripts
python demo_sightrag/sightrag_images.py
python demo_sightrag/sightrag_video.py
python demo_sightrag/sightrag_livecam.py

# Colab
Upload notebooks/SightRAG_v0.2_Demo.ipynb → Run All

# Unit tests
python -m pytest tests/ -v

Usage

Image Folder

rag = SightRAG()
rag.index("./shelf_photos/")
results = rag.query("find empty shelf")
rag.show(results)

Video File

rag = SightRAG()
rag.index("./cctv_footage.mp4")
results = rag.query("person near exit door")
rag.show(results, save="./evidence/")

Mixed Folder (images + videos)

rag = SightRAG()
rag.index("./my_data/")

Live Camera

rag = SightRAG()
rag.index(source="camera")
results = rag.query("find person")

Reference Image Query

results = rag.query(reference="./sample_shelf.jpg")
rag.show(results)

Custom Domain

rag = SightRAG(domain_hint="pcb defect solder joint")
rag.index("./circuit_boards/")
results = rag.query("find defective solder joint")

Visualize Results

results = rag.query("find person")

# Display on screen
rag.show(results)

# Save annotated images with bounding boxes
rag.show(results, save="./output/")

OCR Search : Read Text (NEW in v0.4)

SightRAG reads text on images during indexing. Query matches both visual features AND text content.

rag = SightRAG(ocr=True)
rag.index("./store_photos/")
results = rag.query("find Calgon")       # matches OCR text on packaging
results = rag.query("find EXIT sign")    # matches text on signs
results = rag.query("find plate AB1234") # matches license plate text

OCR runs at INDEX time only. Zero overhead at query time.

Multimodal Understanding (NEW in v0.4)

Optional LLM-powered semantic understanding. Runs on top candidates only, not all images.

# Local model (free, private)
rag = SightRAG(ocr=True, multimodal="qwen2-vl")
results = rag.query("find damaged product", understand=True)

# API model (most accurate)
rag = SightRAG(multimodal="gpt-4o", api_key="...")
results = rag.query("find suspicious activity", understand=True)

# Without understand=True, query is instant (no LLM call)
results = rag.query("find person")  # fast, no LLM

Grounding DINO : Any Domain

Detect ANY object by text description. No training needed.

rag = SightRAG(detector="grounding-dino")
rag.index("./circuit_boards/")
results = rag.query("find cracked solder joint")
rag.show(results)

Works on any domain: medical, industrial, satellite, retail : just type what to find.

Person Re-ID : Cross-Camera Tracking

Track same person across multiple cameras.

rag = SightRAG(embedder="reid")
rag.index("./camera_01/")
rag.index("./camera_02/")
results = rag.query(reference="./suspect.jpg")
rag.show(results)

Returns every camera and timestamp where that person appeared.

CLI Tool

pip install sightrag[cli]

sightrag index ./photos/
sightrag index ./video.mp4 --fps 2
sightrag query "find person near exit"
sightrag query --reference ./suspect.jpg --top-k 10
sightrag show --query "find person" --save ./output/
sightrag status
sightrag clear
sightrag serve --port 8000

Re-ranking : Better Accuracy

Improves result quality on large similar datasets.

rag = SightRAG(rerank=True)
rag.index("./100k_shelf_images/")
results = rag.query("find empty shelf", top_k=5)
# Fetches top 100, re-ranks to best 5

Pluggable Models

Custom Detector

from sightrag import SightRAG
from sightrag.detectors.base import DetectorBase

class MyDetector(DetectorBase):
    def __init__(self):
        self.model = load_my_model()

    def detect(self, image, confidence=0.25):
        preds = self.model.predict(image)
        return [
            {"bbox": [p.x1, p.y1, p.x2, p.y2],
             "label": p.label,
             "confidence": p.score,
             "crop": image.crop((p.x1, p.y1, p.x2, p.y2))}
            for p in preds if p.score >= confidence
        ]

rag = SightRAG(detector=MyDetector())

Custom Embedder

from sightrag.embedders.base import EmbedderBase
import numpy as np

class MyEmbedder(EmbedderBase):
    embed_dim = 768

    def embed_image(self, image):
        vec = self.model.encode_image(image)
        return vec / np.linalg.norm(vec)

    def embed_text(self, text, domain_hint=None):
        vec = self.model.encode_text(text)
        return vec / np.linalg.norm(vec)

rag = SightRAG(embedder=MyEmbedder())

Custom Store

from sightrag.store.base import VectorStoreBase

class MyStore(VectorStoreBase):
    def add(self, id, embedding, metadata): ...
    def search(self, query_vector, top_k): ...
    def count(self): ...
    def clear(self): ...

rag = SightRAG(store=MyStore())

Speed : Auto Backend Selection

SightRAG automatically picks the fastest available backend:

No configuration needed. SightRAG auto-detects what's installed and picks the fastest.

REST API

pip install sightrag[api]
sightrag-server

Result Format

{
    "image_path":  "./photos/shelf_042.jpg",
    "score":       0.9134,
    "label":       "bottle",
    "confidence":  0.8721,
    "bbox":        [120, 45, 380, 290],
    "timestamp":   "",
    "source_type": "image"
}

Storage

rag = SightRAG(store="sqlite")    # default, up to 100k
rag = SightRAG(store="chroma")    # medium scale
rag = SightRAG(store="qdrant")    # production, 1M+

Architecture

Docker

docker-compose up

API at http://localhost:8000/docs

Three Library Ecosystem

Roadmap

License

Apache 2.0

Author

Built by Venkatkumar Rajan

• GitHub: https://github.com/VK-Ant
• LinkedIn: https://linkedin.com/in/vk-ant
• Portfolio: https://vk-ant.github.io/Venkatkumar