Update app.py

app.py

@@ -16,33 +16,43 @@ client = OpenAI(
 
 FINAL_NOTE = "Responses are generated from retrieved hadith evidence and the system is still under improvement."
 
+
 SYSTEM_PROMPT = """
 You are Hadithi AI, a professional assistant for explaining retrieved hadith evidence.
 
-The user's message already includes the retrieved hadith evidence from the API.
-Your job is to explain that evidence clearly, naturally, and faithfully.
+The user message contains:
+1) a question or topic
+2) retrieved hadith evidence from the API
 
-You must base your answer only on the hadiths provided in the user's message.
-Do not invent extra hadiths, extra sources, or unsupported claims.
+Your task:
+- Read the retrieved hadiths carefully
+- Explain only what is supported by the retrieved hadiths
+- Do not invent extra hadiths or unsupported claims
+- Do not produce a generic bullet summary
+- Do not produce the old flat style like:
+  "The hadiths provided emphasize..."
+  followed by bullet points
 
 STRICT OUTPUT FORMAT:
-Write the final answer in exactly this order:
+You must output exactly these parts and in this order:
 
 Answer:
-Write exactly one polished explanatory paragraph in English unless the user clearly asks for Arabic.
-The paragraph must:
-- begin naturally in a style like: "The retrieved hadiths show that..."
-- explain the meaning of the retrieved hadiths in a smooth, human, article-like way
-- mention short Arabic phrases from the hadith only when useful
-- include the meaning of the Arabic phrase naturally in the sentence
-- avoid bullets
-- avoid robotic summary language
-- avoid repeating the same point
-- avoid quoting full hadiths
-- stay grounded in the actual retrieved evidence only
+Write one single polished paragraph in natural English.
+This paragraph must:
+- begin in a smooth explanatory style similar to:
+  "The retrieved hadiths show that..."
+- explain the meaning of the hadith evidence clearly
+- sound human, thoughtful, and elegant
+- include short Arabic phrases only when useful
+- place Arabic phrases naturally inside the explanation, with meaning
+- not use bullet points
+- not sound robotic
+- not repeat the same idea
+- not quote long hadith text
+- summarize the evidence faithfully
 
 Hadith Evidence:
-After the paragraph, list the hadiths provided by the user.
+Then list all retrieved hadiths in a clean way.
 For each hadith, use exactly this structure:
 
 - Source: ...
@@ -51,89 +61,248 @@ For each hadith, use exactly this structure:
 - Text: ...
 
 FINAL LINE:
-End with this exact sentence:
+End with this exact line:
 Responses are generated from retrieved hadith evidence and the system is still under improvement.
 
-IMPORTANT RULES:
-- Do not create extra headings such as Main Insight, Key meanings, or What the Hadiths Show.
-- Do not separate Arabic and English into different blocks in the Answer section.
-- Keep full Arabic hadith text only in the Hadith Evidence section.
-- If the evidence only partially answers the question, say so clearly.
-- Be elegant, clear, modern, and trustworthy.
-
-EXAMPLE STYLE FOR THE ANSWER SECTION:
-The retrieved hadiths show that mercy (raḥma / الرحمة) in Islam is both a divine attribute and a quality believers should live by. In one hadith, the Prophet ﷺ teaches a duʿā’ that ends with “forgive me … and have mercy on me” (“فاغفر لي … وارحمني”), showing that a Muslim constantly needs Allah’s mercy in addition to forgiveness. Another hadith connects mercy with healing through the supplication “place Your mercy on earth” (“فاجعل رحمتك في الأرض”), which presents mercy as something that brings relief and cure. The Prophet ﷺ also said, “Mercy is not removed except from one who is truly wretched” (“لا تنزع الرحمة إلا من شقي”), which shows that mercy is a sign of goodness in the heart, while losing it reflects spiritual hardness. The clearest summary appears in the hadith, “The merciful are shown mercy by the Most Merciful” (“الراحمون يرحمهم الرحمن”), teaching that those who show mercy to people receive mercy from Allah. Together, these hadiths present mercy as something to seek from Allah, to embody in the heart, and to extend to others.
-""".strip()
+FORBIDDEN:
+- Do not create headings like:
+  Main Insight
+  What the Hadiths Show
+  Key meanings
+  Supporting evidence summary
+- Do not start with bullet points
+- Do not write a short outline before the paragraph
+- Do not say "The hadiths provided emphasize..." and then list bullets
+- Do not skip the Answer paragraph
+
+EXAMPLE OF THE DESIRED ANSWER STYLE:
+Answer:
+The retrieved hadiths show that mercy (raḥma / الرحمة) in Islam is both something believers ask from Allah and something they hope to experience in healing, forgiveness, and worship. One hadith includes the supplication "place Your mercy on earth" ("فاجعل رحمتك في الأرض"), presenting mercy as a source of relief and cure, while another links moments of worship with pausing at verses of mercy and praying for it, which shows that mercy is not only a theological idea but also a lived spiritual practice. Together, these narrations present mercy as divine care that the believer actively seeks in prayer, illness, and devotion.
 
+Hadith Evidence:
+- Source: Example
+- Grade: Example
+- Why it matters: Example
+- Text: Example
 
-def is_arabic_request(text: str) -> bool:
-    if not text:
-        return False
-    return bool(re.search(r'[\u0600-\u06FF]', text))
+Responses are generated from retrieved hadith evidence and the system is still under improvement.
+""".strip()
 
 
 def normalize_quotes(text: str) -> str:
     if not text:
         return ""
-    text = text.replace("“", '"').replace("”", '"')
-    text = text.replace("‘", "'").replace("’", "'")
-    return text
+    return (
+        text.replace("“", '"')
+            .replace("”", '"')
+            .replace("‘", "'")
+            .replace("’", "'")
+    )
+
+
+def extract_answer_and_evidence(raw_text: str):
+    """
+    Try to split model output into Answer and Hadith Evidence.
+    If headings are missing, do a best-effort fallback.
+    """
+    text = raw_text.strip()
+
+    # Normalize common variants
+    text = re.sub(r"(?im)^#+\s*answer\s*:?\s*$", "Answer:", text)
+    text = re.sub(r"(?im)^#+\s*hadith evidence\s*:?\s*$", "Hadith Evidence:", text)
+    text = text.replace("Supporting Hadiths:", "Hadith Evidence:")
+    text = text.replace("Referenced Hadiths:", "Hadith Evidence:")
+    text = text.replace("Evidence:", "Hadith Evidence:")
 
+    answer_match = re.search(r"(?is)Answer:\s*(.*?)(?:\n\s*Hadith Evidence:|\Z)", text)
+    evidence_match = re.search(r"(?is)Hadith Evidence:\s*(.*)$", text)
 
-def clean_answer(text: str, user_message: str = "") -> str:
-    if not text:
-        return ""
+    answer = answer_match.group(1).strip() if answer_match else ""
+    evidence = evidence_match.group(1).strip() if evidence_match else ""
 
-    text = normalize_quotes(text.strip())
-
-    replacements = {
-        "Answer Insight:": "Answer:",
-        "Short answer:": "Answer:",
-        "Main Insight:": "Answer:",
-        "Key Finding:": "Answer:",
-        "Supporting Hadiths:": "Hadith Evidence:",
-        "Referenced Hadiths:": "Hadith Evidence:",
-        "What the Hadiths Show:": "Hadith Evidence:",
-        "What the Hadiths Cover:": "Hadith Evidence:",
-        "Evidence:": "Hadith Evidence:",
-        "Hadiths:": "Hadith Evidence:",
-        "Closing Note:": FINAL_NOTE,
-        "Note:": FINAL_NOTE,
-    }
-
-    for old, new in replacements.items():
-        text = text.replace(old, new)
-
-    # Normalize headings if model outputs markdown headings
-    text = re.sub(r"(?im)^#+\s*answer\s*:?\s*$", "Answer:", text)
-    text = re.sub(r"(?im)^#+\s*hadith evidence\s*:?\s*$", "Hadith Evidence:", text)
+    if not answer and not evidence:
+        return text.strip(), ""
+
+    return answer, evidence
+
+
+def clean_intro_bullets(text: str) -> str:
+    """
+    Remove old-style bullet summaries at the beginning of the answer if the model adds them.
+    """
+    lines = [line.rstrip() for line in text.splitlines()]
+    cleaned = []
+    bullet_phase = True
+
+    for line in lines:
+        stripped = line.strip()
+
+        if bullet_phase and (
+            stripped.startswith("- ")
+            or re.match(r"^\d+\.\s+", stripped)
+            or stripped.lower().startswith("the hadiths provided")
+        ):
+            continue
+
+        if stripped:
+            bullet_phase = False
+
+        cleaned.append(line)
+
+    result = "\n".join(cleaned).strip()
+    return result
+
+
+def clean_answer_paragraph(answer: str) -> str:
+    answer = normalize_quotes(answer)
+    answer = clean_intro_bullets(answer)
+
+    # remove obvious unwanted headings inside answer
+    bad_headings = [
+        "Main Insight:",
+        "What the Hadiths Show:",
+        "Key meanings:",
+        "Supporting evidence summary:",
+        "Short answer:",
+    ]
+    for h in bad_headings:
+        answer = answer.replace(h, "")
+
+    # remove leftover bullets inside answer start
+    answer = re.sub(r"(?m)^\s*-\s+", "", answer)
+    answer = re.sub(r"\n{2,}", "\n\n", answer).strip()
+
+    # force single paragraph
+    answer = re.sub(r"\s*\n\s*", " ", answer)
+    answer = re.sub(r"\s{2,}", " ", answer).strip()
+
+    # if model starts badly, nudge it
+    if answer and not answer.lower().startswith("the retrieved hadiths show"):
+        answer = "The retrieved hadiths show that " + answer[0].lower() + answer[1:] if len(answer) > 1 else "The retrieved hadiths show that " + answer
 
-    # Remove numbering before headings
-    text = re.sub(r"(?im)^\s*\d+\.\s*(Answer:)", r"\1", text)
-    text = re.sub(r"(?im)^\s*\d+\.\s*(Hadith Evidence:)", r"\1", text)
+    return answer
+
+
+def parse_hadith_blocks_from_user_message(user_message: str):
+    """
+    Fallback parser:
+    Extract hadith entries directly from the user message if the model fails
+    to produce a proper Hadith Evidence section.
+    """
+    lines = user_message.splitlines()
+    blocks = []
+    i = 0
+
+    source_line_pattern = re.compile(r'^[A-Za-z0-9_\-]+(?:\s+[A-Za-z0-9_\-]+)*\s+#\d+', re.IGNORECASE)
+
+    while i < len(lines):
+        line = lines[i].strip()
+
+        if source_line_pattern.match(line):
+            source = line
+            grade = ""
+            text_lines = []
+            score = ""
+
+            if i + 1 < len(lines) and lines[i + 1].strip().lower().startswith("grade:"):
+                grade = lines[i + 1].strip()
+                # keep score if included on same line
+                i += 2
+            else:
+                i += 1
+
+            while i < len(lines):
+                current = lines[i].strip()
+                if source_line_pattern.match(current):
+                    break
+                if current:
+                    text_lines.append(current)
+                i += 1
+
+            full_text = " ".join(text_lines).strip()
+
+            why = infer_why_it_matters(full_text)
+
+            blocks.append({
+                "source": source,
+                "grade": grade if grade else "Grade: Unknown grade",
+                "why": why,
+                "text": full_text if full_text else "[No text provided]",
+            })
+        else:
+            i += 1
+
+    return blocks
+
+
+def infer_why_it_matters(hadith_text: str) -> str:
+    t = hadith_text
 
-    # Remove duplicate final note before re-adding once
+    if "ارحم" in t or "رحمتك" in t or "رحمة" in t:
+        if "شفاء" in t or "اشف" in t or "الوجع" in t:
+            return "Connects mercy with healing, relief, and supplication."
+        return "Directly relates to mercy as a theme in prayer or belief."
+
+    if "شفاء" in t or "اشف" in t:
+        return "Relates to healing and seeking Allah’s cure."
+
+    if "آية رحمة" in t or "باية رحمة" in t:
+        return "Shows how verses of mercy were treated in worship and recitation."
+
+    if "غفر" in t or "اغفر" in t:
+        return "Relates to forgiveness and seeking Allah’s pardon."
+
+    return "Provides supporting context connected to the retrieved topic."
+
+
+def format_hadith_evidence(blocks) -> str:
+    if not blocks:
+        return "- Source: Not available\n- Grade: Not available\n- Why it matters: The retrieved evidence could not be formatted automatically.\n- Text: Not available"
+
+    formatted = []
+    for b in blocks:
+        formatted.append(
+            f"- Source: {b['source']}\n"
+            f"- Grade: {b['grade']}\n"
+            f"- Why it matters: {b['why']}\n"
+            f"- Text: {b['text']}"
+        )
+    return "\n\n".join(formatted)
+
+
+def clean_answer(model_text: str, user_message: str) -> str:
+    text = normalize_quotes(model_text.strip())
+
+    # Remove final note if model already included it; we add once at the end
     text = re.sub(rf"(?s)\n*{re.escape(FINAL_NOTE)}\s*$", "", text).strip()
 
-    # Reduce excessive blank lines
-    text = re.sub(r"\n{3,}", "\n\n", text).strip()
+    answer, evidence = extract_answer_and_evidence(text)
+    answer = clean_answer_paragraph(answer if answer else text)
 
-    # Ensure Answer section exists
-    if "Answer:" not in text:
-        text = "Answer:\n" + text
+    # If model failed to give usable evidence, build it from the user message
+    if not evidence or "- Source:" not in evidence:
+        parsed_blocks = parse_hadith_blocks_from_user_message(user_message)
+        evidence = format_hadith_evidence(parsed_blocks)
+    else:
+        evidence = normalize_quotes(evidence).strip()
 
-    # Ensure Hadith Evidence section exists
-    if "Hadith Evidence:" not in text:
-        text += "\n\nHadith Evidence:\n"
+    final = f"Answer:\n{answer}\n\nHadith Evidence:\n{evidence}\n\n{FINAL_NOTE}"
+    final = re.sub(r"\n{3,}", "\n\n", final).strip()
+    return final
 
-    # If Arabic requested, lightly relabel only headings, keep body as model returned
-    if is_arabic_request(user_message):
-        text = text.replace("Answer:", "الجواب:")
-        text = text.replace("Hadith Evidence:", "الأحاديث المسترجعة:")
 
-    # Add final note once
-    text = text.rstrip() + "\n\n" + FINAL_NOTE
-    return text
+def build_user_message(message: str) -> str:
+    """
+    Wrap the incoming API text so the model better understands that the user's
+    message contains both the question/topic and retrieved hadith evidence.
+    """
+    return f"""User request and retrieved hadith evidence are below.
+
+Please answer using only this evidence.
+
+Retrieved content:
+{message}
+""".strip()
 
 
 def chat(message, history):
@@ -141,21 +310,22 @@ def chat(message, history):
 
     for user_msg, assistant_msg in history:
         if user_msg:
-            messages.append({"role": "user", "content": user_msg})
+            messages.append({"role": "user", "content": build_user_message(user_msg)})
         if assistant_msg:
             messages.append({"role": "assistant", "content": assistant_msg})
 
-    messages.append({"role": "user", "content": message})
+    wrapped_message = build_user_message(message)
+    messages.append({"role": "user", "content": wrapped_message})
 
     try:
         response = client.chat.completions.create(
             model=MODEL_ID,
             messages=messages,
-            temperature=0.1,
-            max_tokens=1100,
+            temperature=0.08,
+            max_tokens=1200,
         )
-        answer = response.choices[0].message.content.strip()
-        answer = clean_answer(answer, user_message=message)
+        model_text = response.choices[0].message.content.strip()
+        answer = clean_answer(model_text, user_message=message)
     except Exception as e:
         answer = f"Error: {str(e)}"