Document Q&A

import json

# How many document passages to retrieve per question (tunable at generation).
RAG_TOP_K = [[[RAG_TOP_K|4]]]


async def _ingest_document(source, text):
    """(JS-callable) Chunk + embed + index one document's text.

    Underscore-prefixed so it is never exposed to the LLM as a tool.
    Returns the number of chunks stored, as a string (for the UI).
    """
    count = await agentop_rag.add_document(source, text)
    return str(count)


def _extract_pdf_text(pdf_bytes):
    """(JS-callable) Extract plain text from a PDF's raw bytes via pypdf."""
    import io
    from pypdf import PdfReader

    reader = PdfReader(io.BytesIO(pdf_bytes))
    return "\n\n".join((page.extract_text() or "") for page in reader.pages)


async def process_user_query(query):
    """RAG-first answering: retrieve the most relevant passages from the uploaded
    document(s), then answer grounded ONLY in them.

    Overrides the default router so retrieval is deterministic (small local
    models are unreliable at deciding to call a search tool themselves).
    """
    hits = await agentop_rag.search(query, RAG_TOP_K)
    if not hits:
        context = "(No document has been uploaded yet, or nothing relevant was found.)"
    else:
        context = "\n\n".join(
            f"[{i + 1}] (from {h['source']}) {h['text']}" for i, h in enumerate(hits)
        )
    grounded_prompt = (
        "Answer the QUESTION using ONLY the CONTEXT passages below. "
        "If the answer is not in the context, say you could not find it in the "
        "document. Cite passage numbers like [1] where relevant.\n\n"
        f"CONTEXT:\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", "")
    )