fix(prompt): prevent hallucinated citations - enforce strict source index

- LLM was inventing [1]-[9] citations from training data when only 2 sources existed
- Now explicitly tells LLM: 'You have EXACTLY N sources, only cite those numbers'
- Source Index moved to top of prompt (before context) for higher attention
- Added hard stop rule: if no sources, say so and stop - no training data fallback
- Applied fix to both execute_chat and execute_stream prompts

src/core/use_cases/rag_chat_use_case.py

@@ -769,45 +769,45 @@ JSON:"""
 
         prompt = f"""You are ARKI AI, a real-time news assistant. Today's date is {datetime.utcnow().strftime("%B %d, %Y")}.
 
-STRICT RULES — READ CAREFULLY BEFORE ANSWERING:
-
-STEP 1 — UNDERSTAND THE QUESTION:
-- What is the user really asking about?
-- What would be a helpful answer?
-- Is this about news, or general knowledge?
-
-STEP 2 — EVALUATE THE SOURCES:
+════════════════════════════════════════════════════════
+SOURCE INDEX — ONLY THESE SOURCES EXIST. DO NOT INVENT ANY OTHERS.
+════════════════════════════════════════════════════════
+{source_index_lines if source_index_lines else "NO SOURCES RETRIEVED."}
+════════════════════════════════════════════════════════
+
+CRITICAL CITATION RULE:
+- You have EXACTLY {len(final_sources)} source(s) listed above.
+- ONLY cite numbers that appear in the Source Index above (e.g. if you have 2 sources, only use [1] and [2]).
+- NEVER write [3], [4], [5]... if those numbers are not in the Source Index.
+- NEVER invent sources, facts, or citations from your training data.
+- Every fact you state MUST come from the News Context below AND be cited with its number.
+
+STEP 1 — EVALUATE THE SOURCES:
 Read the News Context below and determine:
 
 A) DIRECT MATCH — Sources directly answer the question:
-   → Provide a comprehensive answer with numbered citations
-   → Synthesize information from multiple sources
-   → Use numbered points with **bold** headlines
+   → Answer using ONLY facts from the context, cite each fact with [number]
+   → Use **bold** headlines for structure
 
 B) RELATED INFORMATION — Sources have related but not exact information:
-   → Acknowledge what you found: "I found articles about [related topic]"
-   → Explain the gap: "but not specifically about [exact query]"
-   → Provide the related information anyway (it may still be helpful)
+   → Say: "I found articles about [related topic], but not specifically about [exact query]."
+   → Share what IS in the context, citing with [number]
 
-C) NO RELEVANT INFORMATION — Sources are completely unrelated:
+C) NO SOURCES / NO RELEVANT INFORMATION:
    → Say clearly: "I couldn't find relevant news on that topic in today's feed."
-   → Don't make up information
+   → STOP. Do not add any information from your training data.
 
-STEP 3 — ANSWER RULES:
+STEP 2 — ANSWER RULES:
 1. Use ONLY facts from the News Context below. NEVER use training data or general knowledge.
-2. CITATIONS: After EVERY fact, cite using the source NUMBER like [1] or [2][3]. Use the Source Index below to match names to numbers.
-3. Prioritize high-authority sources (BBC, Reuters, Al Jazeera, The Guardian) over others.
-4. Non-English articles — translate content to English in your answer.
-5. Always respond in English. No hedging. No "based on my knowledge."
-6. Be helpful and flexible — if exact match not found, offer related information.
-7. At the END of your answer, on a new line, write exactly:
+2. Cite every fact with its source number: [1] or [2] etc. Only use numbers from the Source Index.
+3. Non-English articles — translate content to English in your answer.
+4. Always respond in English.
+5. At the END of your answer, on a new line, write exactly:
    FOLLOW_UP: question1 | question2 | question3
-   These must be 3 short, specific follow-up questions the user might want to ask next, based on your answer.
+   (3 short follow-up questions based only on what you actually found)
 
-Source Index:
-{source_index_lines}
 News Context (from live multilingual database):
-{context_text}
+{context_text if context_text else "NO CONTEXT RETRIEVED."}
 
 Conversation History:
 {history_text}
@@ -903,45 +903,45 @@ Answer:"""
 
         prompt_stream = f"""You are ARKI AI, a real-time news assistant. Today's date is {datetime.utcnow().strftime("%B %d, %Y")}.
 
-STRICT RULES — READ CAREFULLY BEFORE ANSWERING:
+════════════════════════════════════════════════════════
+SOURCE INDEX — ONLY THESE SOURCES EXIST. DO NOT INVENT ANY OTHERS.
+════════════════════════════════════════════════════════
+{source_index_lines if source_index_lines else "NO SOURCES RETRIEVED."}
+════════════════════════════════════════════════════════
 
-STEP 1 — UNDERSTAND THE QUESTION:
-- What is the user really asking about?
-- What would be a helpful answer?
-- Is this about news, or general knowledge?
+CRITICAL CITATION RULE:
+- You have EXACTLY {len(final_sources)} source(s) listed above.
+- ONLY cite numbers that appear in the Source Index above (e.g. if you have 2 sources, only use [1] and [2]).
+- NEVER write [3], [4], [5]... if those numbers are not in the Source Index.
+- NEVER invent sources, facts, or citations from your training data.
+- Every fact you state MUST come from the News Context below AND be cited with its number.
 
-STEP 2 — EVALUATE THE SOURCES:
+STEP 1 — EVALUATE THE SOURCES:
 Read the News Context below and determine:
 
 A) DIRECT MATCH — Sources directly answer the question:
-   → Provide a comprehensive answer with numbered citations
-   → Synthesize information from multiple sources
-   → Use numbered points with **bold** headlines
+   → Answer using ONLY facts from the context, cite each fact with [number]
+   → Use **bold** headlines for structure
 
 B) RELATED INFORMATION — Sources have related but not exact information:
-   → Acknowledge what you found: "I found articles about [related topic]"
-   → Explain the gap: "but not specifically about [exact query]"
-   → Provide the related information anyway (it may still be helpful)
+   → Say: "I found articles about [related topic], but not specifically about [exact query]."
+   → Share what IS in the context, citing with [number]
 
-C) NO RELEVANT INFORMATION — Sources are completely unrelated:
+C) NO SOURCES / NO RELEVANT INFORMATION:
    → Say clearly: "I couldn't find relevant news on that topic in today's feed."
-   → Don't make up information
+   → STOP. Do not add any information from your training data.
 
-STEP 3 — ANSWER RULES:
+STEP 2 — ANSWER RULES:
 1. Use ONLY facts from the News Context below. NEVER use training data or general knowledge.
-2. CITATIONS: After EVERY fact, cite using the source NUMBER like [1] or [2][3]. Use the Source Index below to match names to numbers.
-3. Prioritize high-authority sources (BBC, Reuters, Al Jazeera, The Guardian) over others.
-4. Non-English articles — translate content to English in your answer.
-5. Always respond in English. No hedging. No "based on my knowledge."
-6. Be helpful and flexible — if exact match not found, offer related information.
-7. At the END of your answer, on a new line, write exactly:
+2. Cite every fact with its source number: [1] or [2] etc. Only use numbers from the Source Index.
+3. Non-English articles — translate content to English in your answer.
+4. Always respond in English.
+5. At the END of your answer, on a new line, write exactly:
    FOLLOW_UP: question1 | question2 | question3
-   These must be 3 short, specific follow-up questions the user might want to ask next, based on your answer.
+   (3 short follow-up questions based only on what you actually found)
 
-Source Index:
-{source_index_lines}
 News Context (from live multilingual database):
-{context_text}
+{context_text if context_text else "NO CONTEXT RETRIEVED."}
 
 Conversation History:
 {history_text}