txtai — All-in-One Embeddings Database

What it is

txtai is a Python library that provides an all-in-one embeddings database for semantic search, retrieval-augmented generation, and LLM workflow orchestration. It combines vector search with SQL queries, letting you search by meaning and filter by metadata in a single call.

It targets developers and data teams who want semantic search without running separate vector databases, embedding services, and orchestration frameworks. txtai is Apache 2.0 licensed and runs locally.

How it saves time or tokens

A typical RAG setup requires a vector database (Pinecone, Weaviate), an embedding model (OpenAI, Sentence Transformers), and an orchestration layer (LangChain). txtai bundles all three into one library. Install it, create an index, and run semantic search in 3 lines of Python.

The estimated token cost for describing a txtai workflow is approximately 500 tokens.

How to use

1. Install txtai:

pip install txtai

1. Create a semantic search index:

from txtai import Embeddings

embeddings = Embeddings()
embeddings.index([
    'AI is transforming search',
    'Vector databases enable semantic queries',
    'RAG combines retrieval with generation'
])

results = embeddings.search('How does AI improve search?', 2)
print(results)
# Returns the most semantically similar entries

1. Combine with SQL for filtered semantic search:

embeddings = Embeddings(content=True)
embeddings.index([
    {'text': 'Python guide', 'category': 'tutorial'},
    {'text': 'Rust performance', 'category': 'benchmark'}
])

results = embeddings.search(
    "SELECT text, score FROM txtai WHERE similar('performance') AND category='benchmark'"
)

Example

from txtai import Embeddings, LLM

# RAG pipeline in txtai
embeddings = Embeddings(content=True)
embeddings.index(documents)  # Your document list

llm = LLM('meta-llama/Llama-3-8b')

# Search + generate
context = embeddings.search('What is RAG?', 3)
prompt = f'Based on this context: {context}\n\nAnswer: What is RAG?'
answer = llm(prompt)
print(answer)

Related on TokRepo

• AI Tools for RAG -- Compare RAG frameworks and retrieval tools
• AI Tools for Research -- Semantic search tools for research workflows

Common pitfalls

• txtai uses Sentence Transformers by default for embeddings. The first run downloads a model, which can be 400MB+. Pre-download the model in CI/CD pipelines to avoid slow first runs.
• SQL integration requires content=True when creating the Embeddings instance. Without it, only index positions are stored, not the actual text content.
• For large datasets (millions of documents), txtai's in-memory index may exceed available RAM. Use the sqlite or postgres backend for persistent storage.