Knowledge Synthesizer

1import sys2import json3import traceback4from gais import Gais567def main():8    try:9        raw = json.loads(sys.stdin.read())10        inputs = raw.get("inputs", {})1112        query = inputs.get("query", "")13        context = inputs.get("context", [])14        model = inputs.get("llmModel", "qwen3.5-35b")15        language = inputs.get("language", "Italian")1617        # Normalize context to list (fan-in gives a list, single connection gives a string)18        if isinstance(context, str):19            context = [context]2021        if not query:22            print(json.dumps({"text": "No query provided."}))23            return2425        if not context or all(not c for c in context):26            print(json.dumps({"text": "No context chunks available. Connect Document Chunk nodes to provide source material."}))27            return2829        # Build context section from chunks30        context_text = ""31        for i, chunk in enumerate(context, 1):32            if chunk:33                context_text += f"--- Source {i} ---\n{chunk}\n\n"3435        # Build prompt (aligned with RAG agent citation and grounding rules)36        system_prompt = (37            f"You are a Knowledge Synthesis assistant. Your job: answer questions using ONLY the provided source materials.\n\n"38            f"## Rules\n"39            f"1. **Never use general knowledge** — Only use information from the sources below.\n"40            f"2. **Always respond in {language}** — Match the response language exactly.\n"41            f"3. **Never fabricate** — If the sources don't contain enough information, say so clearly.\n"42            f"4. **Cite every factual claim** using the format: [Source: <chunk_id>]\n"43            f"   - The Chunk ID is in each source's header: [Document Path: ... | Chunk ID: <id> | Relevance: ...]\n"44            f"   - Multiple sources: [Source: id1, id2]\n"45            f"   - Place citation IMMEDIATELY after each claim.\n"46            f"5. **Prioritize high-relevance sources** — Sources with higher Relevance scores are more likely to be relevant.\n"47            f"6. **Structure your answer** — Use headings, bullet points, or numbered lists when appropriate for clarity.\n"48        )4950        user_prompt = f"Question: {query}\n\nSource Materials:\n\n{context_text}"5152        messages = [53            {"role": "system", "content": system_prompt},54            {"role": "user", "content": user_prompt},55        ]5657        result = Gais.llm.chat(messages, model=model)58        print(json.dumps({"text": result.text}))5960    except Exception as e:61        error_output = {62            "error": str(e),63            "errorType": type(e).__name__,64            "traceback": traceback.format_exc(),65        }66        print(json.dumps(error_output), file=sys.stderr)67        sys.exit(1)686970if __name__ == "__main__":71    main()