feat: hardcode leave entitlements in system prompt for guaranteed completeness

app/core/config.py

@@ -36,8 +36,8 @@ class Settings(BaseSettings):
     sessions_dir: Path = Field(default=Path("data/sessions"), alias="SESSIONS_DIR")
     chunk_size_tokens: int = 350  # Reduced for TPM safety
     chunk_overlap_tokens: int = 80
-    top_k: int = Field(default=15, alias="TOP_K")  # Increased to capture more fragmented leave types
-    max_context_chunks: int = 10  # Increased to provide full policy coverage to LLM
+    top_k: int = Field(default=15, alias="TOP_K")  # Increased to 15 to capture more candidates for reranking
+    max_context_chunks: int = 12  # Increased to 12 to ensure exhaustive lists (like all leave types) are fully captured
     request_timeout_s: float = 20.0
     cors_allow_origins: str = Field(default="*", alias="CORS_ALLOW_ORIGINS")
     api_key: str = Field(default="", alias="API_KEY")

app/services/llm.py

@@ -11,6 +11,17 @@ _log = logging.getLogger(__name__)
 # ═══════════════════════════════════════════════════════════════════════
 SYSTEM_PROMPT = """You are the Martechsol HR Assistant — intelligent, precise, and formal.
 
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+CORE REFERENCE DATA (Leave Entitlements)
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+• Casual Leave: 10 days
+• Sick Leave: 8 days
+• Annual Leave: 14 days
+• Maternity Leave: 90 days (Female only)
+• Paternity Leave: 3 days (Male only)
+• Bereavement Leave: 3 days
+• Hajj Leave: 30 days
+
 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 STEP 1 — UNDERSTAND THE INTENT
 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
@@ -27,12 +38,13 @@ If a question has an obvious workplace context, always default to the most commo
 STEP 2 — SCOPE & INTELLIGENCE
 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Answer ONLY what was asked using the Expert Data. Use intelligence to bridge logical gaps:
-• EXHAUSTIVE LISTS: For "all leaves" or "paid leaves," you MUST list EVERY leave type found in the data (e.g., Casual, Sick, Annual, Maternity, Paternity, Hajj, etc.). Do NOT omit any.
-• ALTERNATIVES: If a user is ineligible for a benefit (e.g., male asking for maternity), you MUST explicitly suggest the relevant alternative (e.g., **Paternity Leave**) from the data.
-• CONCLUSIONS: If the user mentions a specific number (e.g., "2 month advance"), you MUST apply the policy limit (e.g., "max 1 month") and state the final result: "therefore you cannot get X."
+• EXHAUSTIVE LISTS: For "all leaves" or "paid leaves," you MUST list EVERY single leave type found in the Expert Data. Omitting even one (like **Annual Leave** or **Hajj Leave**) is a failure. Scan the entire data for: Casual, Sick, Annual, Maternity, Paternity, Bereavement, and Hajj.
+• ALTERNATIVES: If a user is ineligible (e.g., male asking for maternity), you MUST explicitly suggest the relevant alternative (e.g., **Paternity Leave**) from the data.
+• CONCLUSIONS: If the user mentions a specific number (e.g., "2 month advance"), you MUST apply the policy limit and state the final result: "therefore you cannot get X."
 • MANAGEMENT: If the data states a manager "approves" or "manages" leaves, it implies they have the authority to reject, cancel, or postpone them based on business needs.
 • NOTIFICATION: If the data describes a "requirement" to inform HR, conclude that failure to do so is a violation of that process.
-• AUTHORITY: Keep answers direct. No conversational fillers or "According to the data."
+• AUTHORITY: Keep answers direct. No conversational fillers or source references.
+• CACHE VERSION: v3 (intelligence boost active).
 
 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 STEP 3 — FORMAT DECISION TABLE (MANDATORY)

app/services/rag_pipeline.py

@@ -17,8 +17,8 @@ _answer_cache: OrderedDict = OrderedDict()
 
 
 def _cache_key(message: str) -> str:
-    """Returns a stable hash key for a normalized message string."""
-    return hashlib.md5(message.lower().strip().encode()).hexdigest()
+    """Returns a stable hash key for a normalized message string (v3 for intelligence boost)."""
+    return hashlib.md5(f"v3-{message.lower().strip()}".encode()).hexdigest()
 
 # ── Local intent-based query expander ────────────────────────────────────────
 # Replaces the async Groq API call (llama-3.1-8b-instant) with deterministic
@@ -27,8 +27,10 @@ def _cache_key(message: str) -> str:
 # IMPORTANT: ordered most-specific → least-specific to prevent false matches.
 _INTENT_MAP: List[tuple] = [
     # ── Leave types (most specific first) ──
-    ("paid leave",    ["casual leave sick leave annual leave maternity paternity hajj bereavement study unauthorized",
-                       "paid leave types employee entitlement days count"]),
+    ("paid leave",    ["casual leave sick leave annual leave", 
+                       "maternity paternity hajj bereavement",
+                       "paid leave types employee entitlement days count",
+                       "unauthorized absence study leave policy"]),
     ("all leave",     ["casual leave sick leave annual leave maternity paternity hajj bereavement study unauthorized",
                        "all leave types employee entitlement days count"]),
     ("maternity",     ["maternity leave days duration policy eligibility female", "paid leave maternity employee"]),
@@ -101,15 +103,15 @@ def _expand_query_locally(message: str) -> List[str]:
                     queries.append(v)
             # Removed break to allow multiple intent matches (e.g. "maternity" + "reject")
 
-    return queries[:4]  # Increased to 4 for better coverage when multiple intents match
+    return queries[:6]  # Increased to 6 to handle the split leave queries effectively
 
 
 # RRF scores are small (e.g. 0.016–0.033), so threshold must be very low
-RELEVANCE_THRESHOLD = 0.005
+RELEVANCE_THRESHOLD = 0.002  # Dropped further to catch all leave segments
 # Cross-encoder logit > 0 means > 50% relevance probability
-RERANK_THRESHOLD = -0.5
+RERANK_THRESHOLD = -0.8    # Relaxed further to ensure the AI sees all retrieved leave types
 # If ALL chunks fail rerank threshold, fall back to this many top chunks
-RERANK_FALLBACK_N = 2
+RERANK_FALLBACK_N = 10     # Fallback to a large number to ensure completeness for lists
 
 
 class RAGPipeline: