Semantic Search

# How many passages to return per search.
SEARCH_TOP_K = [[[SEARCH_TOP_K|5]]]


async def _ingest(source, text):
    """(JS-callable) Index text as individual passages, one per non-empty line.

    Per-line indexing keeps each note / FAQ entry / list item its own searchable
    unit (rather than merging a short blob into one chunk), so ranking is
    meaningful. Underscore-prefixed so it is never exposed to the LLM as a tool.
    Returns the number of passages stored, as a string (for the UI).
    """
    passages = [line.strip() for line in text.splitlines() if line.strip()]
    total = 0
    for passage in passages:
        total += await agentop_rag.add_document(source, passage)
    return str(total)


async def process_user_query(query):
    """Return the most semantically similar passages — no LLM generation.

    Retrieval is deterministic and runs entirely on the embedding model, so this
    template never loads a chat model.
    """
    hits = await agentop_rag.search(query, SEARCH_TOP_K)
    if not hits:
        return "No matches yet — index some text first, then search."
    lines = []
    for i, h in enumerate(hits):
        score = round(float(h["score"]) * 100)
        lines.append(f"[{i + 1}] {score}% match · {h['source']}\n{h['text']}")
    return "\n\n".join(lines)