Inference Conversation History (Beta)

Telnyx Inference now supports Conversation History, the ability to persist, index, and semantically search call transcripts across your Telnyx communications.

Because Telnyx owns the full stack from network to GPU, we can embed and index conversations at the point of generation, with no external vector database, no pipeline glue, and no data leaving the private backbone. The call happens on our SIP infrastructure, the transcript gets embedded on our Inference infrastructure, and the search runs through the same API, one system, one vendor, one contract.

This matters for any team building AI agents on top of real-time communications. Instead of stitching together a telephony provider, a transcription service, a vector database, and an inference API, you enable conversation_persistence on your SIP connection and search via GET /v2/ai/conversation_histories. The results come back chunked and relevance-scored, ready to use as RAG context for your agent.

How it works

Enable conversation_persistence on a SIP connection (ip, credential, or fqdn). When a call completes, Telnyx automatically embeds the transcript, chunks it, and indexes it. You then search across all persisted conversations with a semantic query, Inference embeds your search text and returns matching chunks with relevance scores, record metadata, and filter support by region, record type, and more.

30 days of indexed retention is included. Extended retention and additional storage options are available for workloads that need longer recall windows.

What's available today

Voice (SIP) is supported at launch. Messaging and Email data sources are on the roadmap, same Inference memory layer, more conversation types to search across.

Pricing

Embed and persist costs $0.0015 per 1K characters (includes 30-day retention). Storage beyond 30 days is $0.60/GiB-month. Vector search is $0.003 per search, with the first 10,000 searches per month free.

Learn more in the Conversation History documentation.