""" Stage 1 — Intent Parser Model: DeepSeek R1 (reasoning model, ideal for structured analysis) Input: raw user prompt Output: structured intent dict used by the code generator """ import json import re from openai import OpenAI from config import CRUSOE_API_KEY, CRUSOE_BASE_URL, INTENT_MODEL client = OpenAI(api_key=CRUSOE_API_KEY, base_url=CRUSOE_BASE_URL) SYSTEM_PROMPT = """You are an expert at analysing AI demo requests and mapping them to the right app template. Given a user's demo idea, output a single JSON object with these fields: REQUIRED (all templates): - template_type: one of "chatbot" | "comparison" | "dashboard" | "form_wizard" - title: short, catchy demo title (max 8 words) - description: one-line description shown below the title - system_prompt: detailed system prompt for the AI assistant inside the demo - model: best Crusoe model for this demo: • "deepseek-ai/DeepSeek-R1" — reasoning, analysis, step-by-step thinking • "moonshotai/Kimi-K2-Instruct" — coding, long documents, structured output • "Qwen/Qwen3-235B-A22B" — general purpose, fast responses, multilingual - features: list of 2-4 key features to highlight CONDITIONAL (chatbot only): - chat_placeholder: placeholder text shown in the chat input box CONDITIONAL (comparison only): - model_a: first model ID (from the list above) - model_a_label: friendly display name for model A - model_b: second model ID - model_b_label: friendly display name for model B CONDITIONAL (dashboard only): - input_label: label for the main input text area - input_placeholder: placeholder text for that input CONDITIONAL (form_wizard only): - steps: list of 3-5 objects, each {"key": "snake_case_name", "question": "Question text?"} Template selection guide: - chatbot: conversational Q&A, support bots, advisors - comparison: show two models side-by-side on the same prompt - dashboard: analyze / summarize pasted text, data, or documents - form_wizard: multi-step intake flows that end with AI recommendations Output ONLY the JSON object. No markdown, no explanations.""" def parse_intent(prompt: str) -> dict: response = client.chat.completions.create( model=INTENT_MODEL, messages=[ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": f"Create a demo for: {prompt}"}, ], temperature=0.3, max_tokens=2000, ) content = response.choices[0].message.content or "" # Strip DeepSeek R1 chain-of-thought tags content = re.sub(r".*?", "", content, flags=re.DOTALL).strip() # Extract the JSON block if surrounded by markdown fences fenced = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", content, re.DOTALL) if fenced: content = fenced.group(1) else: # Grab the outermost { ... } brace = re.search(r"\{.*\}", content, re.DOTALL) if brace: content = brace.group() intent = json.loads(content) return _apply_defaults(intent) def _apply_defaults(intent: dict) -> dict: intent.setdefault("template_type", "chatbot") intent.setdefault("title", "AI Demo") intent.setdefault("description", "Powered by Crusoe Managed Inference") intent.setdefault("system_prompt", "You are a helpful AI assistant.") intent.setdefault("model", "Qwen/Qwen3-235B-A22B") intent.setdefault("features", []) # Chatbot defaults intent.setdefault("chat_placeholder", "Ask me anything...") # Comparison defaults intent.setdefault("model_a", "deepseek-ai/DeepSeek-R1") intent.setdefault("model_a_label", "DeepSeek R1") intent.setdefault("model_b", "moonshotai/Kimi-K2-Instruct") intent.setdefault("model_b_label", "Kimi K2") # Dashboard defaults intent.setdefault("input_label", "Paste your content here:") intent.setdefault("input_placeholder", "Enter text to analyze...") # Form wizard defaults intent.setdefault("steps", [ {"key": "requirement", "question": "What is your main requirement?"}, {"key": "context", "question": "Can you describe your use case in more detail?"}, {"key": "constraints", "question": "Are there any constraints or preferences to keep in mind?"}, ]) return intent