Results using txtai's benchmark script. All vectors generated using all-MiniLM-L6-v2
Index and search times are similar for all methods. For larger sources (like FiQA), the index time will be lower as IVF training is not required.
| Method | Disk (MB) | NDCG_10 |
|---|---|---|
| Faiss IVF | 13.7 MB | 0.4761 |
| Faiss IVF SQ4 | 2.3 MB | 0.4739 |
| from unsloth import FastModel | |
| from datasets import load_dataset | |
| from txtai.pipeline import HFTrainer | |
| # Load training dataset | |
| train = load_dataset(...) | |
| # Full fine-tune (16-bit) | |
| # Set load_in_4bit=True and trainer.lora=True to train a QLoRA model |
| from txtai import Embeddings | |
| # SQLite + NumPy | |
| embeddings = Embeddings(content=True, backend="numpy") | |
| embeddings.index(["test"]) | |
| embeddings.save("test") | |
| # SQLite File | |
| # sqlite3 test/documents "SELECT * FROM sections" | |
| # 0|0|test||2026-01-16 16:12:03.410194+00:00 |
| # Install TxtAI with Torch CPU | |
| pip install txtai torch==2.9.1+cpu -f https://download.pytorch.org/whl/torch | |
| # Disk usage is much less | |
| # Torch CPU = 1.1G | |
| # Torch CUDA = 6.9G | |
| # Number of installed packages is also less | |
| # Torch CPU = 27 | |
| # Torch CUDA = 43 |
| # | |
| # pip install ai-edge-torch txtai | |
| # | |
| # See https://github.com/google-ai-edge/ai-edge-torch | |
| import torch | |
| import ai_edge_torch | |
| import numpy as np |
| from txtai import Embeddings | |
| from txtai.pipeline import HFOnnx | |
| # Export as quantized onnx model | |
| # Model Size = 1.1 MB! | |
| path = HFOnnx()( | |
| "neuml/biomedbert-hash-nano-embeddings", | |
| "pooling", "model.onnx", True | |
| ) |
| from txtai import Embeddings | |
| # Load the 100K most viewed Wikipedia articles | |
| embeddings = Embeddings() | |
| embeddings.load(provider="huggingface-hub", container="neuml/txtai-wikipedia-slim") | |
| # Show a random article | |
| embeddings.search(""" | |
| SELECT id, text, 'https://en.wikipedia.org/wiki/' || replace(id, ' ', '_') url | |
| FROM txtai |
| from txtai import Agent | |
| # Define tools | |
| tools = [ | |
| "websearch", # Runs a websearch using default engine | |
| "webview", # Loads a web page | |
| ] | |
| # Define LLM | |
| model = "Qwen/Qwen3-4B-Instruct-2507" |
| import numpy as np | |
| from txtai.ann import ANNFactory | |
| # Index 10M vectors using llama.cpp style quants | |
| ann = None | |
| for _ in range(1000): | |
| # Generate batch of vectors | |
| batch = np.random.rand(10000, 768).astype(np.float32) | |
| if not ann: |
| # RAG Quick Start | |
| # Easy to use way to get started with RAG using YOUR data | |
| # | |
| # For a complete application see this: https://github.com/neuml/rag | |
| # | |
| # TxtAI has 70+ example notebooks covering everything the framework provides | |
| # Examples: https://neuml.github.io/txtai/examples | |
| # | |
| # Install TxtAI | |
| # pip install txtai[pipeline-data] |