vaad

vaad is the value guard between speech-to-text and your tools for Hindi/Hinglish agents. It never lets a wrong amount, time or date execute silently: it is sure, or it asks. It parses spoken-form values ("dedh hazaar" = 1500, "saade aath baje", lakh/crore, quantities, phone numbers, tense-aware "kal"), with a span, a calibrated confidence and an audit trail, and works as a library, a CLI, a Pipecat processor (pip install "vaad[pipecat]"), a LiveKit Agents mixin (pip install "vaad[livekit]") or an OpenAI-compatible gateway. See docs/POSITIONING.md.

vaad guard "kal saade aath baje dedh lakh rupaye bhej do" --ref-date 2026-11-09
# use: amount 150000   confirm: "₹1,50,000, sahi hai?"   ask: "'saade aath baje': subah ya shaam?"

Live demo against a local model (Qwen3.5-4B, synthetic requests; python examples/demo_payment.py, transcript):

User: 9876501234 pe paanch sau rupaye bhej do
  without vaad: send_money(amount_inr=50000)      <- wrong in 4 of 4 runs
  with vaad:    "₹500, sahi hai?"
User: 98123 45670 ko dedh hazaar transfer kar do
  without vaad: send_money(amount_inr=1000)       <- wrong in 4 of 4 runs
  with vaad:    "1,500, sahi hai?"

This compares the bare model with the guard's output; it is not an end-to-end run of the guard in front of the model.

from datetime import date
from vaad.canon import annotate_turn

annotate_turn("mera order kal tak bhej do, dedh hazaar rupaye", date(2026, 10, 24))
# 'mera order kal [=2026-10-25] tak bhej do, dedh hazaar [=1500] rupaye'

What is measured (details and caveats in docs/RESULTS.md, calibration in docs/CALIBRATION.md):

• On 189 held-out human-typed tool-calling tasks, the normalizer raised pass^4 by 16 to 22 points on three small models (all intervals above 0) and cut wrong values from 23-42% to 7-9% with canon.mode: both.
• Outputs the parser is at least 90% sure of were right 97-100% of the time on held-out data.
• These tasks were written to contain values. In real chat such values are rare, and spoken input through ASR is not tested yet (planned for v2, see docs/AUDIO_SPEC.md).
• Lexicons are AI-written and not yet native-reviewed.

Status: pre-alpha. See docs/ for the plan, results, methods, data, architecture and decisions.

Development

Requires uv.

uv sync
uv run pytest                     # tests
uv run ruff check . && uv run ruff format --check .
uv run mypy                       # strict on canon/, router/, grammar/
uv run python -m harness.leakage  # gate 10
uv run vaad serve --config config/vaad.yaml   # copy from config/vaad.example.yaml

Layout

License

Apache-2.0.