Function rcli_rag_query

src/api/rcli_api.cpp:2551–2637

}

const char* rcli_rag_query(RCLIHandle handle, const char* query) {
    if (!handle || !query) return "";
    auto* engine = static_cast<RCLIEngine*>(handle);
    if (!engine->rag_ready || !engine->rag_retriever || !engine->rag_embedding) {
        return "{\"error\": \"RAG not initialized. Run rcli_rag_ingest() first.\"}";
    }

    std::lock_guard<std::mutex> lock(engine->mutex);

    auto query_embedding = engine->rag_embedding->embed(std::string(query));
    if (query_embedding.empty()) {
        return "{\"error\": \"Failed to embed query\"}";
    }

    auto results = engine->rag_retriever->retrieve(
        std::string(query), query_embedding, 5);

    if (results.empty()) {
        engine->last_rag_result = "{\"results\": [], \"answer\": \"No relevant documents found.\"}";
        return engine->last_rag_result.c_str();
    }

    // Build context from retrieved chunks
    std::string context;
    for (size_t i = 0; i < results.size(); i++) {
        context += "[" + std::to_string(i + 1) + "] " + std::string(results[i].text) + "\n\n";
    }

    // Generate answer using LLM with retrieved context.
    // The system prompt enforces speech-friendly output since responses are
    // streamed to TTS — no bullets, numbered lists, markdown, or citations.
    std::string rag_system = rastack::apply_personality(
        "You are RCLI, a voice assistant that answers questions using provided documents. "
        "Your response will be spoken aloud by a text-to-speech engine, so you MUST follow these rules:\n"
        "- Answer in plain, natural conversational sentences only.\n"
        "- NEVER use bullet points, numbered lists, markdown, or any formatting.\n"
        "- NEVER include citations like [1], [2], or source references.\n"
        "- NEVER use parentheses for asides or abbreviations.\n"
        "- Do NOT use <think> tags or internal reasoning.\n"
        "- Keep your answer concise — 2 to 4 sentences maximum.\n"
        "- If the answer is not in the context, say so briefly.",
        engine->personality_key);

    std::string rag_prompt =
        "Context:\n" + context + "\n"
        "Question: " + std::string(query);

    std::string answer;
    if (engine->pipeline.using_metalrt()) {
        auto& mrt = engine->pipeline.metalrt_llm();
        std::string rag_full = mrt.profile().build_chat_prompt(rag_system, {}, rag_prompt);
        answer = mrt.generate_raw(rag_full, nullptr);
        engine->metalrt_kv_continuation_len = 0;
    } else if (engine->pipeline.llm().is_initialized()) {
        answer = engine->pipeline.llm().generate(
            engine->pipeline.llm().build_chat_prompt(rag_system, {}, rag_prompt),
            nullptr);
    } else {