NOTICE

streamRAG-audio
Copyright 2026 ehzawad

A single-A5000 cascaded emulation of the *Model-Triggered Streaming RAG* control
policy from the Stream RAG paper (Arora et al., Meta + CMU, arXiv:2510.02044),
applied to spoken CRAG queries. It reuses this author's own prior work:
  - StreamRAG        https://github.com/ehzawad/streamRAG        (typed streaming RAG)
  - streamrag-local  https://github.com/ehzawad/streamrag-local  (fully-local variant; this repo's base)
  - omni-voice-lab   https://github.com/ehzawad/omni-voice-lab   (Silero-VAD turn/barge-in logic, demo only)
The MIT license above covers the source code in this repository.

What this is NOT (honest scope — see docs/AUDIO.md):
  - NOT a reproduction of the paper's end-to-end speech-in/speech-out model, its
    audio-conditioned trigger, its joint response post-training, or its web+KG tools.
  - This reproduces the paper's modality-agnostic *control policy* over a cascade
    (Qwen3-TTS offline query synthesis + faster-whisper ASR + local Qwen3.5-9B).
    The reused coordinator revalidates evidence at commit, so it is
    SAFER-BUT-DIFFERENT from the paper (which trusts the latest tool result); it is
    not a bit-exact copy. (An earlier single-voice Chatterbox synthesis was the
    prior baseline; it is retained only as frozen evidence, not a live component.)

Third-party components (fetched or run locally; NOT redistributed here):
  - Qwen3.5-9B                    Apache-2.0        (Alibaba / Qwen)
  - bge-large-en-v1.5             MIT               (BAAI)
  - llama.cpp                     MIT               (ggml-org)
  - Silero VAD                    MIT               (snakers4)         [turn/barge-in; retired demo only]
  - faster-whisper / Whisper      MIT               (SYSTRAN / OpenAI) [offline full-clip ASR]
  - Qwen3-TTS (CustomVoice)       Apache-2.0        (Alibaba / Qwen)   [9-voice query synthesis, current]
  - Chatterbox-TTS                MIT               (Resemble AI)      [prior single-voice baseline; retired, frozen result only]
  - PydanticAI, FastAPI, Qdrant, etc. — see pyproject.toml / requirements-audio.txt

Evaluation data — CRAG (CC BY-NC 4.0, NON-COMMERCIAL):
  - CRAG — Comprehensive RAG Benchmark (Meta), arXiv:2406.04744,
    https://github.com/facebookresearch/CRAG
  - This repository DOES include a small CRAG-derived evaluation subset under
    data/crag_eval/ (10 sealed test + 5 dev questions, gold answers, and 250
    curated source pages), inherited from streamrag-local and curated from the
    official CRAG Task-1/2 development release. It is provided for NON-COMMERCIAL
    RESEARCH use only, consistent with CRAG's CC BY-NC 4.0 license and with
    attribution to Meta.
  - Any spoken audio synthesized from these questions (the current Qwen3-TTS
    9-voice set under data/audio_crag_qwen/) is a derivative of CRAG and is
    therefore likewise NON-COMMERCIAL; it is gitignored, not redistributed, and
    must be regenerated locally. This is a CRAG-style local auto-eval, NOT an official CRAG/KDD
    leaderboard score, and its absolute numbers are NOT comparable to the paper's
    (local Qdrant retrieval is not the paper's 100K-document web+KG pipeline).

Model weights and the full CRAG dataset are not distributed as part of this
repository. No affiliation with or endorsement by the paper's authors, Meta, CMU,
or any upstream project is implied.