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_id = inputs.get("llmModelId", "")15        # 'auto' (default) matches the question's language; locale codes map16        # to names; legacy full-name values from older canvases are accepted17        # via the allowlist. Anything else (wired upstream output, free-form18        # strings) falls back to auto: the value lands in the system prompt,19        # so unknown strings must never pass through.20        language = inputs.get("language") or "auto"21        language_names = {22            "de": "German", "en": "English", "es": "Spanish", "fr": "French",23            "it": "Italian", "ja": "Japanese", "ko": "Korean", "nl": "Dutch",24            "pt": "Portuguese", "ru": "Russian", "zh": "Chinese",25        }26        lang_name = None27        if isinstance(language, str) and language.strip() not in ("", "auto"):28            value = language.strip()29            if value in language_names:30                lang_name = language_names[value]31            elif value in language_names.values():32                lang_name = value33            else:34                print(f"[language] unrecognized value {value!r} ignored - using auto", file=sys.stderr)35        if lang_name:36            language_rule = f"**Always respond in {lang_name}** — Match the response language exactly."37        else:38            language_rule = (39                "**Respond in the language of the question** — Match the "40                "question's language, not the sources' language."41            )4243        # Normalize context to a flat list of strings. Fan-in delivers a raw44        # list whose elements may THEMSELVES be lists (a Text[] source arrives45        # as one nested list) — flatten one level so each chunk is its own46        # numbered Source instead of a python-repr blob.47        if isinstance(context, str):48            context = [context]49        flat_context = []50        for item in context:51            if isinstance(item, list):52                flat_context.extend(str(c) for c in item if c)53            elif item:54                flat_context.append(str(item))55        context = flat_context5657        if not query:58            print(json.dumps({"text": "No query provided."}))59            return6061        if not context or all(not c for c in context):62            print(json.dumps({"text": "No context chunks available. Connect Document Chunk nodes to provide source material."}))63            return6465        # Build context section from chunks66        context_text = ""67        for i, chunk in enumerate(context, 1):68            if chunk:69                context_text += f"--- Source {i} ---\n{chunk}\n\n"7071        # Build prompt (aligned with RAG agent citation and grounding rules)72        system_prompt = (73            f"You are a Knowledge Synthesis assistant. Your job: answer questions using ONLY the provided source materials.\n\n"74            f"## Rules\n"75            f"1. **Never use general knowledge** — Only use information from the sources below.\n"76            f"2. {language_rule}\n"77            f"3. **Never fabricate** — If the sources don't contain enough information, say so clearly.\n"78            f"4. **Cite every factual claim** using the format: [Source: <chunk_id>]\n"79            f"   - The Chunk ID is in each source's header: [Document Path: ... | Chunk ID: <id> | Relevance: ...]\n"80            f"   - Multiple sources: [Source: id1, id2]\n"81            f"   - Place citation IMMEDIATELY after each claim.\n"82            f"5. **Prioritize high-relevance sources** — Sources with higher Relevance scores are more likely to be relevant.\n"83            f"6. **Structure your answer** — Use headings, bullet points, or numbered lists when appropriate for clarity.\n"84            f"7. **Knowledge-graph sources are authoritative for counts and relationships** — A source starting with"85            f" 'Entity profile:' (or 'Fused ranking') comes from the organization's knowledge graph, not a document."86            f" Its per-link-type counts are EXACT (e.g. 'employed_by: 18 - <names>')."87            f" When the question asks how many people/entities hold a relationship (how many employees, who works"88            f" there, how many X does Y regulate), ANSWER WITH THE GRAPH COUNT and list the named entities,"89            f" citing [Source: Knowledge Graph]. State that the number is what the knowledge graph records;"90            f" you may add partial figures from document chunks as supporting detail, but never answer"91            f" 'cannot be determined' when an Entity profile block contains the relevant count.\n"92        )9394        user_prompt = f"Question: {query}\n\nSource Materials:\n\n{context_text}"9596        messages = [97            {"role": "system", "content": system_prompt},98            {"role": "user", "content": user_prompt},99        ]100101        # thinking=False: the reasoning on/off eval (qwen3.5-35b-a3b) showed102        # synthesis quality held with reasoning off — same grounded, cited103        # conclusions at ~8-25x lower latency — so skip the wasted CoT pass.104        result = Gais.llm.chat(messages, model_id=model_id, thinking=False)105        print(json.dumps({"text": result.text}))106107    except Exception as e:108        error_output = {109            "error": str(e),110            "errorType": type(e).__name__,111            "traceback": traceback.format_exc(),112        }113        print(json.dumps(error_output), file=sys.stderr)114        sys.exit(1)115116117if __name__ == "__main__":118    main()