How does Moss achieve sub-10ms retrieval for voice agents?

Moss loads a compact index into memory inside your agent runtime. Queries execute locally as function calls with zero network overhead. End-to-end latency including embedding inference is 3.1ms at p50 and 5.4ms at p99, benchmarked on 100,000 documents.

Which voice agent frameworks does Moss integrate with?

Moss has dedicated integration packages: pipecat-moss for Pipecat (pipeline processor), vapi-moss for VAPI (Custom Knowledge Base webhook), and elevenlabs-moss for ElevenLabs (client tool). LiveKit integrates via the core Python SDK with a context injection pattern. Install the relevant package and follow the integration guide.

Can Moss handle real-time conversation workloads?

Yes. Moss is designed for the real-time loop of voice agents: ASR transcription, context retrieval, LLM generation, and TTS. Retrieval adds less than 10ms to each turn, which is imperceptible in conversation.

Does Moss work offline for voice agents?

Yes. Once an index is loaded, all queries run entirely in-memory on the device with no network required. This makes Moss ideal for voice agents that need to work in low-connectivity environments or on-device.

How does Moss compare to Pinecone or Qdrant for voice agent retrieval?

Moss delivers 3.1ms p50 latency versus 432ms for Pinecone and 597ms for Qdrant. The difference is architectural: Moss runs locally inside your runtime, while Pinecone and Qdrant require network round-trips to cloud infrastructure.

Voice AI

Semantic Search for Voice Agents

From voice agents to AI copilots, Moss gives your product <10ms semantic retrieval without managing vector infrastructure.

Start Free Explore Docs

I need to reschedule my appointment

Sure! I see you're booked with Dr. Chen, Thursday at 2pm. What day works better for you?

Does next Tuesday work?

Yes, Tuesday at 2pm is open. Want me to move it there?

Yes, please.

Rescheduled to Tuesday, 2pm

<10 ms

P99 retrieval latency

0 ms

network round-trips

100x

faster than cloud vector DBs

Your voice agent needs context in under 10 milliseconds.

Network round-trips to cloud vector databases add 300-900ms of dead air. Moss runs retrieval locally, inside your agent runtime, so your agent responds without the pause.

The Problem

Dead air kills voice agents

Traditional RAG sends retrieval over the network to Pinecone or Qdrant — and round-trips mean silence. Users notice gaps as short as 200ms. The LLM was never the bottleneck. Retrieval is.

How Moss Solves This

Local retrieval, zero network hops

Moss runs search inside your agent runtime. No network round-trip to a cloud database. Retrieval completes in under 10ms, locally.

Built for streaming conversation

Designed for the real-time loop of ASR, retrieval, LLM, and TTS. Moss fits into the critical path without adding perceptible latency.

Works with every voice stack

Drop-in integration with LiveKit, Pipecat, VAPI, ElevenLabs, and Hume AI. Install the SDK and start querying in minutes.

Ship real-time retrieval in minutes

Explore docs

from moss import MossClient

client = MossClient(PROJECT_ID, PROJECT_KEY)

docs = [{ "text": "How do I track my order?" }]

await client.add_docs("my-index", docs)

Frequently asked questions

Ready to ship faster AI products?

Moss gives you production-ready semantic retrieval without infrastructure complexity.

Test performance Talk to an Engineer

How Moss Solves This

Local retrieval, zero network hops

Moss runs search inside your agent runtime. No network round-trip to a cloud database. Retrieval completes in under 10ms, locally.

Built for streaming conversation

Designed for the real-time loop of ASR, retrieval, LLM, and TTS. Moss fits into the critical path without adding perceptible latency.

Works with every voice stack

Drop-in integration with LiveKit, Pipecat, VAPI, ElevenLabs, and Hume AI. Install the SDK and start querying in minutes.