💥 BREAKING CHANGE: All metric scores are now HIGHER THE BETTER. Read changelog →

Integrations Overview

DeepEval integrates with the frameworks, model providers, and data stores teams already use to build LLM applications. Use these pages to connect tracing, evaluation, synthetic data, and model configuration to your existing stack.

Frameworks

Framework integrations let DeepEval evaluate entire execution traces without manually instrumenting every intermediate step. They preserve the ordered agent execution needed for trajectory-based evals, while also supporting end-to-end evals and component-level evals across the agents, chains, tools, and workflows already in your stack.

Evaluation Models

Evaluation model integrations configure the LLM provider DeepEval uses for LLM-as-a-judge metrics, synthetic data generation, conversation simulation, and prompt optimization. Pick the provider that matches your infrastructure, latency, privacy, and cost needs.

Speech Models

Speech model integrations configure the text-to-speech and speech-to-text providers DeepEval uses for voice simulation. Every provider apart from OpenAI is reached over plain HTTP, so none of them need a package beyond deepeval itself.

Speech-To-Text (STT)

STT transcribes your agent's spoken replies into the Turn.content your metrics judge, so its accuracy bounds how faithfully every multi-turn metric sees the conversation.

Text-To-Speech (TTS)

TTS speaks the simulated user's messages to your agent, so its clarity is what your agent's own speech recognition has to work with.

Voice Agent Connectors

Voice agent connectors are how DeepEval reaches the agent under test during a voice simulation: they open a live session, play the simulated user's audio, and capture the spoken reply. Pick the one matching where your agent runs.

Custom agents are covered by WebSocketConnector and CallbackVoiceConnector — see Voice Connectors.

Vector DBs

Vector database integrations show how to connect retrieval systems to DeepEval so RAG metrics can evaluate the context your application actually retrieves. Use these examples to benchmark retrieval quality and end-to-end RAG behavior.

Others

Integrations that don't fit cleanly into the categories above — typically training/eval-time hooks rather than runtime tracing.

On this page