$devvkit learn --librarie chromadb-&-txtai-guide

ChromaDB & txtai Guide

[vector-db][embeddings][rag][similarity]

AI / LLM Tools

Install

# ChromaDB:
pip install chromadb
# txtai:
pip install txtai
# Both: python -m pip install sentence-transformers

ChromaDB is the simplest vector database. `import chromadb; chromadb.Client()` gives you an in-memory or persistent (`./chroma_data`) vector store. Add documents with `collection.add` and search with `collection.query`. It handles embedding generation or accepts pre-computed vectors.

txtai goes further — it's an all-in-one embeddings database with built-in RAG, semantic search, LLM-powered extraction, and workflow pipelines. A single `pip install txtai` gives you a full semantic search + QA pipeline. The YAML config defines embeddings, indexing, and queries declaratively.

Both support multiple embedding models: `all-MiniLM-L6-v2` (fast), `text-embedding-3-small` (OpenAI), or local models via sentence-transformers. For production, ChromaDB scales with `clickhouse` or `postgres` backends. GUI: Chroma's optional web UI, or FiftyOne for dataset visualization.

ChromaDB Setup

Start ChromaDB— In-memory or persistent.

import chromadb

# In-memory (for testing):
client = chromadb.Client()

# Persistent:
client = chromadb.PersistentClient(path='./chroma_data')

# Start server (multi-process):
# chroma run --path /db --port 8000
client = chromadb.HttpClient(host='localhost', port=8000)

Chroma HTTP server— Run as a service.

# Terminal:
pip install chromadb
chroma run --path ./chroma_data --port 8000 --host 0.0.0.0

# Docker:
docker run -p 8000:8000 chromadb/chroma

# Python client:
client = chromadb.HttpClient(host='localhost', port=8000)
collection = client.get_or_create_collection('production')

ChromaDB CRUD

Create collection— Add documents with embeddings.

collection = client.create_collection('my_docs')

collection.add(
    documents=[
        'Chroma is a vector database',
        'txtai builds AI-powered search',
        'Embeddings capture semantic meaning'
    ],
    metadatas=[
        {'source': 'wiki', 'topic': 'database'},
        {'source': 'wiki', 'topic': 'ai'},
        {'source': 'blog', 'topic': 'machine-learning'}
    ],
    ids=['doc1', 'doc2', 'doc3']
)

Chroma — update/delete— Maintain collections.

# Update document:
collection.update(
    ids=['doc1'],
    documents=['Updated document text'],
    metadatas=[{'source': 'wiki', 'version': 2}]
)

# Delete:
collection.delete(ids=['doc2'])

# List collections:
client.list_collections()

# Delete collection:
client.delete_collection('my_docs')

Chroma — batch ingest— Efficient large-scale add.

# Batch add 10k documents:
ids = [f'doc{i}' for i in range(10000)]
chunks = [f'Document number {i}' for i in range(10000)]

collection.add(
    documents=chunks,
    ids=ids,
    batch_size=100  # Chroma handles batching
)

# With custom embeddings (for speed):
from sentence_transformers import SentenceTransformer
model = SentenceTransformer('all-MiniLM-L6-v2')
embeddings = model.encode(chunks).tolist()

collection.add(
    embeddings=embeddings,
    documents=chunks,
    ids=ids
)

ChromaDB Search

Query by similarity— Semantic search.

results = collection.query(
    query_texts=['What is a vector database?'],
    n_results=2
)
# Returns: {documents, metadatas, distances, ids}
print(results['documents'][0])

# Filter by metadata:
collection.query(
    query_texts=['AI tools'],
    where={'source': 'wiki'},
    n_results=10
)

txtai

txtai — quick start— Launch txtai embeddings.

from txtai import Embeddings

embeddings = Embeddings(path='sentence-transformers/all-MiniLM-L6-v2')

# Index documents
data = [
    'Chroma is a vector database for AI',
    'txtai provides semantic search and RAG',
    'Vector search finds meaning, not keywords'
]
embeddings.index(data)

# Search
results = embeddings.search('semantic search', limit=2)
for result in results:
    print(f'{result["text"]} — score: {result["score"]}'

RAG Pipeline

txtai RAG pipeline— Full question-answering.

from txtai import Application

# YAML config
app = Application('''
writable: true
embeddings:
  path: sentence-transformers/all-MiniLM-L6-v2

rag:
  path: openai/gpt-4o-mini
  prompt: 'Answer based on context: {context}\nQuestion: {question}'
''')

# Index content
app.add('Chroma is a vector database built for AI')
app.add('txtai provides semantic search and RAG pipelines')
app.index([])

# Ask questions
answer = app.search('What database is used for vector storage?')
print(answer)