Under the Hood

Zero Hallucinations.
Pure Determinism.

Every response is traced back to a real database row or a verified policy rule. Here's how a single voice query flows through the system — from microphone to answer.

The Processing Pipeline

A voice query moves through four stages. At stage three, the system forks depending on whether the query needs live data or a policy lookup.

01

Instant Voice Streaming

LiveKit WebRTC captures speech in real time with sub-50ms network latency. Raw audio is streamed as Opus-encoded packets — no HTTP polling, no waiting for full utterances.

  • Peer-to-peer WebRTC data channels
  • Adaptive bitrate and jitter buffering
  • DTLS-SRTP encrypted transport
LiveKit WebRTC
02

Intent Parsing & Entity Extraction

Transcribed speech is processed through a structured NLU pipeline that extracts intents, entities (order IDs, dates, categories), and confidence scores with zero ambiguity.

  • Slot-filling extraction for IDs, dates, names
  • Multi-turn context preservation
  • Confidence thresholds for ambiguity rejection
NLU Pipeline
The Fork — Step 03
Data-Driven Path

Deterministic Database Lookup

When the intent requires live data, the system executes a parameterized SQL query against your PostgreSQL or internal API. No vector search, no probabilistic retrieval — just exact row lookups.

01 Intent parsed → parameterized query built

02 SELECT * FROM orders WHERE id = $1

03 Result: exact row returned (12ms)

04 Row data injected into response template

Rulebook Path

Strict Policy Verification

When the intent matches a policy domain, the system performs an indexed lookup against your fixed rulebook. Responses are validated against policy rules before delivery — with out-of-scope questions gracefully deflected.

01 Intent parsed → rulebook index scanned

02 MATCH: Rule #R-204 (Electronics, 30-day window)

03 Response drafted from policy template

04 VERIFY: Compliance check ✓ — delivered

Merge — Step 04
04

Natural Voice Synthesis

The verified response — whether sourced from live data or policy rules — is converted to natural speech using a low-latency TTS engine and streamed back over WebRTC. Total pipeline: under 500ms.

  • Streaming TTS with word-level timestamps
  • Prosody matching for natural delivery
  • Result streamed as soon as first chunk is ready
< 500ms E2E

End-to-End Guarantee

Every response is either sourced from a verified database row or a policy rule. There is no generative layer that can hallucinate.

< 500ms

Total latency

0

Hallucinations