Why Raven
General-purpose models like GPT and Claude are trained for broad text-based use cases. They require extensive prompting to work well for customer service — and even then, they can be unreliable. Raven is different: the right conversational behavior is built in. Most models are trained to be good at everything - writing code, creative writing, answering questions on any topic, even generating images. It means:- They’re slower (because they’re huge and handle everything)
- They act like text chatbots (because that’s what most people use them for)
- They need a lot of coaching to work well for phone conversations
Conversation-native
Trained specifically for customer service interactions across voice and chat. Responses are concise and natural without extensive prompting — add raw information to your knowledge base and Raven converts it into conversational responses.
Faster
Sub-300ms median latency with consistent response times. No long-tail latency spikes that disrupt conversation flow.
More accurate
Higher accuracy on PolyAI’s customer service benchmarks than general-purpose models, with fewer errors in function calling and knowledge base grounding.
Multilingual
Supports 24+ languages natively with near-perfect language consistency. Set the response language and Raven speaks it — even with English-only prompts.
Additional capabilities
- Stronger date and time logic — handles relative dates, scheduling scenarios, and format conversions that trip up general-purpose models
- Reliable function calling — trained on real Agent Studio projects, so it calls functions with the right parameters and doesn’t confuse when to respond vs. when to act
- No hallucination — grounded in your knowledge base topics. Raven says “I don’t know” rather than making something up, and is robust to irrelevant retrieved content
- Optimized for Agent Studio — understands topics, flows, and PolyAI’s knowledge retrieval patterns by default
Raven 3.5
Raven 3.5 is the latest and most capable Raven model. It supports both voice and chat, and is recommended for all new deployments.- Auto-reasoning — automatically decides when to think deeper before responding, improving accuracy on complex tasks like date calculations without adding latency on simple turns
- Out-of-domain detection — identifies when a request falls outside the agent’s scope, enabling cleaner handoffs and knowledge gap tracking
- Built-in safety — robust guardrails against misuse, with built-in protection against hallucinations
- Custom style following — respects custom persona and style instructions, including emotion tags for TTS, formatting rules, and channel-specific tone
- 24+ languages — more natural multilingual outputs than V3, with near-perfect language consistency
Raven V3 is a legacy model superseded by Raven 3.5. It supports the same 24 languages but lacks auto-reasoning, out-of-domain detection, chat support, and the safety and style improvements in 3.5. Raven 3.5 outperforms V3 in all scenarios — there is no use case where V3 is the better choice. V3 remains available in Voice configuration for existing deployments that have not yet migrated.
Supported languages
Raven supports the following languages:- All languages
- Strongest performance
Arabic, Bulgarian, Cantonese, Croatian, Czech, Dutch, English, French, German, Greek, Hindi, Hindi (Romanized/Hinglish), Italian, Japanese, Korean, Mandarin (China), Mandarin (Taiwan), Polish, Portuguese (Brazil), Portuguese (Portugal), Serbian, Spanish (US), Swedish, Turkish
Getting started
Select Raven 3.5 in Voice configuration and/or Chat configuration. It is the recommended model for all scenarios — voice, chat, multichannel, and multilingual.Related pages
Model selection
Compare Raven with OpenAI and Amazon Bedrock models.
Training data
Transparency on datasets used to train Raven.
Bring your own model
Connect your own LLM endpoint to PolyAI.