Voice Transcriber

# Accumulated transcripts for this session: [{"source": name, "text": text}]
TRANSCRIPTS = []


async def _transcribe_audio(samples, name):
    """(JS-callable) Transcribe decoded audio samples and remember the result.

    Underscore-prefixed so it is never exposed to the LLM as a tool. ``samples``
    is a Float32Array of mono 16 kHz samples handed over by the page JS.
    Returns the transcript text (for the UI).
    """
    text = (await agentop_ml.transcribe(samples)).strip()
    TRANSCRIPTS.append({"source": name, "text": text})
    return text


async def process_user_query(query):
    """Answer questions / summarise grounded ONLY in the accumulated transcripts.

    Overrides the default router so context is deterministic (small local
    models are unreliable at deciding to call tools themselves).
    """
    if not TRANSCRIPTS:
        return (
            "No audio has been transcribed yet - upload a voice note or "
            "audio file first."
        )
    context = "\n\n".join(
        f"[{i + 1}] (from {t['source']}) {t['text']}"
        for i, t in enumerate(TRANSCRIPTS)
    )
    grounded_prompt = (
        "Answer the QUESTION using ONLY the TRANSCRIPTS below. "
        "If the answer is not in the transcripts, say you could not find it. "
        "Cite transcript numbers like [1] where relevant.\n\n"
        f"TRANSCRIPTS:\n{context}\n\nQUESTION: {query}"
    )
    # TEMPLATE_SYSTEM_PROMPT only exists as a JS global; the query bridge falls
    # back to it when custom_prompt is empty (same idiom as the base template).
    return await process_user_query_wllama(
        grounded_prompt, globals().get("TEMPLATE_SYSTEM_PROMPT", "")
    )