Eliminate first Groq API call: replace LLM query rewrite with instant local keyword expander (~3-6s saved per request)

app/services/rag_pipeline.py

@@ -20,6 +20,82 @@ def _cache_key(message: str) -> str:
     """Returns a stable hash key for a normalized message string."""
     return hashlib.md5(message.lower().strip().encode()).hexdigest()
 
+# ── Local intent-based query expander ────────────────────────────────────────
+# Replaces the async Groq API call (llama-3.1-8b-instant) with deterministic
+# Python rules — runs in microseconds, saves 3–6s per request.
+# Rules mirror the exact intent-detection logic from the old LLM prompt.
+# 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"]),
+    ("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", "paid leave maternity employee"]),
+    ("paternity",     ["paternity leave days duration policy", "paid leave paternity employee"]),
+    ("hajj",          ["hajj leave days duration policy", "paid leave hajj religious"]),
+    ("bereavement",   ["bereavement leave death family days", "compassionate leave policy"]),
+    ("sick leave",    ["sick leave days count policy", "medical leave employee entitlement"]),
+    ("casual leave",  ["casual leave days count policy", "leave types employee"]),
+    ("annual leave",  ["annual leave days count policy", "leave entitlement per year"]),
+    ("study leave",   ["study leave education policy days", "employee study leave entitlement"]),
+    ("leave",         ["casual leave sick leave annual leave maternity paternity hajj bereavement study unauthorized",
+                       "leave types employee entitlement days"]),
+    # ── Office timing ──
+    ("office hour",   ["office working hours schedule", "workday start end time shift"]),
+    ("work hour",     ["office working hours schedule", "workday start end time shift"]),
+    ("timing",        ["office working hours schedule", "workday start end time shift hours"]),
+    ("schedule",      ["office working hours schedule", "workday start end time"]),
+    # ── Salary / Pay ──
+    ("salary",        ["salary structure payroll compensation amount", "monthly pay increment deduction"]),
+    ("pay",           ["salary structure payroll compensation", "payment date schedule"]),
+    ("payroll",       ["salary payroll structure compensation", "monthly pay deduction"]),
+    ("compensation",  ["salary compensation structure payroll", "monthly pay amount"]),
+    # ── Benefits / Allowances ──
+    ("allowance",     ["allowances perks medical bonuses fuel transport reimbursements", "employee benefits"]),
+    ("benefit",       ["allowances perks medical bonuses reimbursements", "employee benefits privileges"]),
+    ("perk",          ["employee perks benefits extras privileges", "allowances bonuses"]),
+    ("fuel",          ["fuel allowance transport reimbursement conveyance petrol"]),
+    ("medical",       ["medical allowance health insurance coverage", "medical benefits employee"]),
+    ("transport",     ["transport allowance fuel reimbursement conveyance"]),
+    ("bonus",         ["bonus performance incentive annual eid festival reward"]),
+    # ── Termination / Resignation ──
+    ("terminat",      ["termination resignation procedure process steps", "notice period exit policy"]),
+    ("resign",        ["resignation procedure steps notice period", "termination exit process"]),
+    ("notice period", ["notice period resignation termination duration days"]),
+    # ── Other HR topics ──
+    ("probat",        ["probation period duration conditions employee"]),
+    ("overtime",      ["overtime compensation extra hours payment policy"]),
+    ("increment",     ["salary increment raise annual review appraisal performance"]),
+    ("appraisal",     ["performance appraisal review increment salary raise"]),
+    ("attendance",    ["attendance policy punctuality late arrival absenteeism"]),
+    ("dress code",    ["dress code uniform attire professional clothing policy"]),
+    ("remote",        ["remote work work from home WFH policy telecommute"]),
+    ("grievance",     ["grievance complaint procedure policy employee rights"]),
+    ("discipline",    ["disciplinary action policy procedure employee"]),
+    ("code of conduct", ["code of conduct policy employee behaviour rules"]),
+]
+
+
+def _expand_query_locally(message: str) -> List[str]:
+    """
+    Expands a user query into targeted search strings using keyword rules.
+    Replaces the async LLM rewrite call — runs in microseconds, zero API cost.
+    Mirrors the exact intent-detection logic previously in the llama-3.1-8b prompt.
+    """
+    msg_lower = message.lower()
+    queries: List[str] = [message]  # Original query always first
+
+    for keyword, variants in _INTENT_MAP:
+        if keyword in msg_lower:
+            for v in variants:
+                if v not in queries:
+                    queries.append(v)
+            break  # Only apply first (most specific) matching intent
+
+    return queries[:3]  # Cap at 3, consistent with previous LLM behaviour
+
+
 # RRF scores are small (e.g. 0.016–0.033), so threshold must be very low
 RELEVANCE_THRESHOLD = 0.01
 # Cross-encoder logit > 0 means > 50% relevance probability
@@ -51,9 +127,9 @@ class RAGPipeline:
             logger.info("Cache HIT for: '%s'", message[:40])
             return _answer_cache[key]
 
-        # ── Step 1: Generate multiple search queries for broad coverage ──
-        queries = await self.llm_service.generate_multi_queries(message, history)
-        logger.info("Generated %d queries for: '%s' → %s", len(queries), message[:40], queries)
+        # ── Step 1: Expand query locally (no API call — instant) ──
+        queries = _expand_query_locally(message)
+        logger.info("Expanded %d queries for: '%s' → %s", len(queries), message[:40], queries)
 
         # ── Step 2: Collect unique chunks across all queries ──
         seen_ids = set()
@@ -81,11 +157,12 @@ class RAGPipeline:
             return {"reply": reply, "retrieved_chunks": []}
 
         # ── Step 4: Deep reranking via Cross-Encoder ──
-        # Enrihc the reranker query with the LLM's expanded search terms
+        # Enrich the reranker query with the first expanded variant (queries[1])
+        # to give the cross-encoder broader semantic context.
         rerank_query = message
-        if len(queries) > 0 and queries[0] != message:
-            rerank_query = f"{message} {queries[0]}"
-            
+        if len(queries) > 1:
+            rerank_query = f"{message} {queries[1]}"
+
         reranked_chunks = self.reranker.rerank(rerank_query, initial_chunks, top_n=self.max_context_chunks)
 
         # Filter by rerank score threshold