Compression Infrastructure
TurboAgents logo

TurboAgents

TurboQuant-style KV-cache and vector compression -integrated into SuperOptiX as both a GEPA vector-store backend and shared RAG retriever.

4 Validated Backends
GEPA Vector Store
Framework-Agnostic
Apache 2.0

SuperOptiX + TurboAgents

TurboAgents integrates into SuperOptiX as both a GEPA vector-store backend and shared RAG retriever, providing compressed retrieval with full validation.

As a GEPA Vector Store

TurboAgents-backed vector stores plug directly into GEPA's optimization pipeline. Use TurboChromaVectorStore, TurboFAISSVectorStore, TurboLanceDBVectorStore, or TurboSurrealDBVectorStore as drop-in replacements for standard vector stores.

Compressed scoring and reranking
Preserves existing GEPA workflows
Programmatic imports from superoptix.optimizers.gepa_rag_adapter

As a RAG Retriever

Use TurboAgents retriever types in SuperOptiX playbooks via YAML RAG blocks. Specify retriever_type as turboagents-chroma, turboagents-faiss, turboagents-lancedb, or turboagents-surrealdb.

Playbook integration via YAML RAG blocks
All four backends validated end-to-end
Compressed retrieval without changing your pipeline

Validated Backends

Every backend has been validated end-to-end inside SuperOptiX

Chroma

Passed
Retriever:turboagents-chroma
Store:TurboChromaVectorStore
chromadb 1.5.5

FAISS

Passed
Retriever:turboagents-faiss
Store:TurboFAISSVectorStore
faiss-cpu

LanceDB

Passed
Retriever:turboagents-lancedb
Store:TurboLanceDBVectorStore
Seeded token validated

SurrealDB

Passed
Retriever:turboagents-surrealdb
Store:TurboSurrealDBVectorStore
OpenAI + Pydantic AI

Installation

Install SuperOptiX with TurboAgents support

1

From PyPI (published package)

uv pip install "superoptix[turboagents]"
2

From source checkout

uv sync --extra turboagents
3

With framework validation

uv sync --extra turboagents --extra frameworks-openai

Configuration

Each backend requires specific configuration. All share common compression parameters: dim, bits, seed, embedding_function, and rerank_top.

FAISS

In-memory compressed retrieval. Requires dim, bits, seed, embedding_function, and rerank_top.

# FAISS - TurboAgents vector store
rag:
  enabled: true
  retriever_type: turboagents-faiss
  config:
    top_k: 5
  vector_store:
    embedding_model: sentence-transformers/all-MiniLM-L6-v2
    embedding_dimension: 64
    bits: 3.5
    seed: 0

Chroma

Adds path and collection_name to the common parameters.

# Chroma - TurboAgents vector store
rag:
  enabled: true
  retriever_type: turboagents-chroma
  config:
    top_k: 5
  vector_store:
    persist_directory: ./.superoptix/turboagents-chroma
    collection_name: documents
    embedding_model: sentence-transformers/all-MiniLM-L6-v2
    embedding_dimension: 64
    bits: 3.5
    seed: 0

LanceDB

Adds uri and table_name for local or remote LanceDB instances.

# LanceDB - TurboAgents vector store
rag:
  enabled: true
  retriever_type: turboagents-lancedb
  config:
    top_k: 5
  vector_store:
    uri: ./.superoptix/turboagents-lancedb
    table_name: documents
    embedding_model: sentence-transformers/all-MiniLM-L6-v2
    embedding_dimension: 64
    bits: 3.5

SurrealDB

Adds url, namespace, database, and table_name for SurrealDB connections.

# SurrealDB - TurboAgents vector store
rag:
  enabled: true
  retriever_type: turboagents-surrealdb
  config:
    top_k: 5
  vector_store:
    url: ws://localhost:8000/rpc
    namespace: test
    database: test
    table_name: documents
    embedding_model: sentence-transformers/all-MiniLM-L6-v2
    embedding_dimension: 64
    bits: 3.5

What TurboAgents Brings to SuperOptiX

Compression infrastructure that sits beneath your existing agent stack

CORE

KV-Cache Compression

Walsh-Hadamard rotation with PolarQuant-style encoding. Compress KV-cache for extended context windows.

CORE

Vector Payload Compression

Reduce vector storage costs and scale retrieval without replacing your existing backend.

HIGH

Compressed Reranking

Apply compressed scoring on retrieved candidates for better precision at lower cost.

HIGH

MLX Native Support

First-class Apple Silicon support with validated 3.5-bit sweet spot on 3B models.

MEDIUM

llama.cpp & vLLM

Runtime wrappers for llama.cpp and experimental vLLM support for server-side inference.

HIGH

Benchmark-First

CLI-driven benchmarks with turboagents bench rag and turboagents bench kv. Measure before you commit.

Current Scope

Honest engineering -here is what the integration supports today and what is still in progress

Working Today

  • Chroma, FAISS, LanceDB, SurrealDB backends validated
  • GEPA vector-store integration
  • Playbook RAG block integration
  • SurrealDB + OpenAI and Pydantic AI paths
  • Programmatic and YAML configuration
  • Dimensions: 64, 128, 256

In Progress

  • Metadata filtering not yet implemented
  • SurrealDB + DSPy path (Ollama/LiteLLM compatibility)
  • SurrealDB wrapper uses sync boundary around async adapter
  • LiteLLM versions 1.82.7–1.82.8 excluded (security)

Resources

Integration Guide

SuperOptiX + TurboAgents setup and configuration

TurboAgents Docs

Standalone TurboAgents documentation

TurboAgents GitHub

Source code, benchmarks, and examples

TurboAgents PyPI

Python package on PyPI

TurboAgents Landing Page

Full product overview on super-agentic.ai

SuperOptiX GitHub

SuperOptiX source with TurboAgents integration

Ready to Add Compressed Retrieval?

Install SuperOptiX with TurboAgents and run compressed retrieval against your existing vector store

uv pip install "superoptix[turboagents]"