fix parsing of prompts

app.py

@@ -6,6 +6,7 @@ Tiny Titan eligible (0.5B parameters).
 
 import os
 import re
+import json
 import gradio as gr
 from huggingface_hub import InferenceClient
 from chemistry import (
@@ -37,6 +38,40 @@ and suggest one simple fix.
 End your response after giving your advice. Do not ask follow-up questions. Do not say 'How do you want to proceed' or similar."""
 
 
+# ---------------------------------------------------------------------------
+# Extraction prompt (LLM -> structured JSON)
+# ---------------------------------------------------------------------------
+EXTRACTION_SYSTEM_PROMPT = """You are an extraction assistant. Your job is to extract ingredients and weights
+from a user's free-form text and return a single JSON object only (no explanation).
+
+Requirements:
+- If the user provided a recipe or ingredient list, return exactly:
+    {"type": "recipe", "ingredients": [{"name": "<ingredient name>", "weight_g": <number>}, ...]}
+- If the user did NOT provide ingredients (they asked a question or for advice), return exactly:
+    {"type": "advice"}
+- Normalise ALL weight formats to grams (g). Accept formats like: "300g", "300 grams",
+    "0.3kg", "10oz", "1 lb", "1lb", "ounces", "kgs", and convert them to grams.
+    Use 1 oz = 28.3495 g, 1 lb = 453.59237 g, 1 kg = 1000 g.
+- Use numeric grams (floats allowed). Round to one decimal place only when necessary.
+- Return only JSON; do not include any trailing text or markdown.
+- Use deterministic output: set sampling temperature to 0.0 (this will be enforced by the caller).
+
+Few-shot examples (input -> JSON):
+
+Input: "300g coconut oil, 200g olive oil, 100g shea butter"
+Output: {"type": "recipe", "ingredients": [{"name": "coconut oil", "weight_g": 300}, {"name": "olive oil", "weight_g": 200}, {"name": "shea butter", "weight_g": 100}]}
+
+Input: "0.3kg olive oil and 10oz coconut oil"
+Output: {"type": "recipe", "ingredients": [{"name": "olive oil", "weight_g": 300}, {"name": "coconut oil", "weight_g": 283.5}]}
+
+Input: "I have 1 lb coconut oil and 250 g olive oil"
+Output: {"type": "recipe", "ingredients": [{"name": "coconut oil", "weight_g": 453.6}, {"name": "olive oil", "weight_g": 250}]}
+
+Input: "How can I make a gentle soap for dry sensitive skin?"
+Output: {"type": "advice"}
+"""
+
+
 # ---------------------------------------------------------------------------
 # AI call
 # ---------------------------------------------------------------------------
@@ -124,6 +159,91 @@ def detect_mode(text: str) -> str:
     return "recipe"
 
 
+# ---------------------------------------------------------------------------
+# LLM extraction and JSON validation
+# ---------------------------------------------------------------------------
+
+def _find_json_in_text(text: str) -> str | None:
+    m = re.search(r"\{.*\}", text, re.DOTALL)
+    return m.group(0) if m else None
+
+
+def _map_ingredients_to_recipe(ingredients: list[dict]) -> dict[str, float] | None:
+    recipe = {}
+    for item in ingredients:
+        name = (item.get("name") or "").lower()
+        weight = item.get("weight_g")
+        try:
+            weight = float(weight)
+        except Exception:
+            return None
+        if weight <= 0:
+            return None
+
+        matched_key = None
+        for alias, key in OIL_ALIASES.items():
+            if alias in name:
+                matched_key = key
+                break
+        if not matched_key:
+            for key in OILS:
+                if key in name:
+                    matched_key = key
+                    break
+
+        if not matched_key:
+            # Could not map ingredient to a known oil
+            return None
+
+        recipe[matched_key] = recipe.get(matched_key, 0.0) + weight
+
+    return recipe
+
+
+def extract_with_llm(user_input: str) -> tuple[str, dict | None]:
+    """Call the LLM to extract ingredients as JSON. Returns (type, recipe_or_none).
+    type is 'recipe' or 'advice'. If extraction fails or JSON is invalid, returns ('advice', None).
+    """
+    try:
+        response = client.chat_completion(
+            model=MODEL,
+            messages=[
+                {"role": "system", "content": EXTRACTION_SYSTEM_PROMPT},
+                {"role": "user", "content": user_input},
+            ],
+            max_tokens=400,
+            temperature=0.0,
+        )
+        content = response.choices[0].message.content
+    except Exception:
+        return ("advice", None)
+
+    # Try to extract JSON from the model output
+    json_text = None
+    try:
+        json_text = _find_json_in_text(content) or content
+        parsed = json.loads(json_text)
+    except Exception:
+        return ("advice", None)
+
+    if not isinstance(parsed, dict) or "type" not in parsed:
+        return ("advice", None)
+
+    if parsed["type"] == "advice":
+        return ("advice", None)
+
+    if parsed["type"] == "recipe":
+        ingredients = parsed.get("ingredients")
+        if not isinstance(ingredients, list):
+            return ("advice", None)
+        mapped = _map_ingredients_to_recipe(ingredients)
+        if not mapped:
+            return ("advice", None)
+        return ("recipe", mapped)
+
+    return ("advice", None)
+
+
 # ---------------------------------------------------------------------------
 # Mode handlers
 # ---------------------------------------------------------------------------
@@ -602,7 +722,15 @@ Your water should not meaningfully interfere with lather or leave extra mineral
 
 
 def run_soap_lab(user_input: str, vegan_only: bool, superfat: float = 5.0, water_hardness: str = "Soft") -> tuple[str, str, str, str]:
-    parsed_recipe = parse_recipe(user_input)
+    # First, ask the LLM to extract a structured recipe JSON (temperature=0.0)
+    extraction_type, extracted = extract_with_llm(user_input)
+    if extraction_type == "advice":
+        # Extraction determined this is an advice request or extraction failed.
+        advice_context = "The user did not provide a parseable recipe. Provide concise, practical soap-making advice based on their question."
+        advice = ask_model(user_input, advice_context)
+        return "", "", advice, water_hardness_recommendation(water_hardness, None)
+
+    parsed_recipe = extracted or {}
     total_oil_weight = sum(parsed_recipe.values()) if parsed_recipe else None
     water_panel = water_hardness_recommendation(water_hardness, total_oil_weight)