refactor: improve keyword extraction and re-ranking prompts

Apply llm-findings.md recommendations:

Keyword extraction (temp 0.0):
- Add negative constraints: "Do NOT invent terms"
- Simplify output format
- Remove verbose example

LLM re-ranking (temp 0.0):
- Add explicit rules section
- Negative constraints: "Do NOT consider document length"
- Clearer output format specification

Both prompts now use temperature=0.0 for deterministic,
consistent outputs.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

src/services/hybrid_rag_service.py

@@ -195,25 +195,21 @@ class HybridRAGService:
         Returns:
             Dictionary with keywords, entities, synonyms, expansions
         """
-        prompt = f"""Extract search terms from this query. For each important word, provide synonyms and expansions.
+        prompt = f"""Extract search terms from this query.
 
 Query: "{query}"
 
-Return ONLY valid JSON:
-{{
-  "core_keywords": ["key", "words", "from", "query"],
-  "synonyms": {{
-    "word": ["alternative", "terms"]
-  }}
-}}
+RULES:
+- Extract ONLY keywords explicitly present or directly implied in the query
+- Do NOT invent terms, concepts, or synonyms not clearly related
+- Do NOT add general knowledge or associations
+- Provide synonyms ONLY for technical terms with well-known alternatives
+- Return valid JSON only, no commentary
 
-Example for "Docker container hosting":
+Return format:
 {{
-  "core_keywords": ["docker", "container", "hosting"],
-  "synonyms": {{
-    "docker": ["containerization", "container runtime"],
-    "hosting": ["server", "infrastructure"]
-  }}
+  "core_keywords": ["words", "from", "query"],
+  "synonyms": {{"term": ["direct", "alternatives"]}}
 }}
 
 JSON:"""
@@ -221,7 +217,8 @@ JSON:"""
         try:
             response = await self.ollama.generate_text(
                 prompt=prompt,
-                model=self.reranker_model
+                model=self.reranker_model,
+                temperature=0.0  # Deterministic for consistent extraction
             )
 
             # Parse JSON response (handle potential extra text)
@@ -623,21 +620,27 @@ JSON:"""
                 for i, r in enumerate(results)
             ])
 
-            prompt = f"""Given this search query and documents, rank them by relevance.
+            prompt = f"""Rank these documents by relevance to the query.
 
 Query: {query}
 
 Documents:
 {docs_text}
 
-Return only the numbers in order of relevance (most relevant first).
-Example: 3,1,5,2,4
+RULES:
+- Rank ONLY by how well content answers the query
+- Do NOT consider document length, formatting, or style
+- Do NOT add explanation or commentary
+- Return ONLY comma-separated numbers, most relevant first
+
+Example output: 3,1,5,2,4
 
 Ranking:"""
 
             response = await self.ollama.generate_text(
                 prompt=prompt,
-                model=self.reranker_model
+                model=self.reranker_model,
+                temperature=0.0  # Deterministic for consistent rankings
             )
 
             # Parse response: "3,1,5,2,4" → [2, 0, 4, 1, 3] (0-indexed)