feat: Replace Gemini with Groq Cloud Llama-3.3 ultra-fast LLM API integration

.env.example

@@ -1,11 +1,11 @@
 # Sarvam AI STT API Key (Saaras v3)
 SARVAM_API_KEY=your_sarvam_api_key_here
 
-# Primary LLM Generation API Settings (Gemini Flash / OpenAI-compatible adapter)
-GEMINI_API_KEY=your_gemini_api_key_here
-LLM_API_KEY=your_gemini_api_key_here
-LLM_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai/
-LLM_MODEL=gemini-2.5-flash
+# Primary LLM Generation API Settings (Groq Cloud Llama-3.3 / OpenAI-compatible adapter)
+GROQ_API_KEY=your_groq_api_key_here
+LLM_API_KEY=your_groq_api_key_here
+LLM_BASE_URL=https://api.groq.com/openai/v1
+LLM_MODEL=llama-3.3-70b-versatile
 
 # Cerebras High-Speed Backup Provider
 CEREBRAS_API_KEY=your_cerebras_api_key_here
@@ -20,9 +20,9 @@ ADAPTION_API_KEY=your_adaption_api_key_here
 EMBEDDING_MODEL_NAME=intfloat/multilingual-e5-small
 
 # Storage & Index paths
-DATA_DIR=./data
-INDEX_DIR=./data/indexes
-PROCESSED_DIR=./data/processed
+DATA_DIR=./knowledge_base
+INDEX_DIR=./knowledge_base/indexes
+PROCESSED_DIR=./knowledge_base/processed
 
 # Server Configuration
 HOST=0.0.0.0

config.py

@@ -191,11 +191,11 @@ DYNAMIC_SEMANTIC_CACHE_ENABLED = os.getenv("DYNAMIC_SEMANTIC_CACHE_ENABLED", "tr
 DYNAMIC_SEMANTIC_CACHE_MAX_ENTRIES = int(os.getenv("DYNAMIC_SEMANTIC_CACHE_MAX_ENTRIES", "2048"))
 DYNAMIC_SEMANTIC_CACHE_THRESHOLD = float(os.getenv("DYNAMIC_SEMANTIC_CACHE_THRESHOLD", "0.92"))
 
-# Tier-1 Primary: Gemini Flash / OpenAI-compatible (High-speed instruction model, ~250ms)
-GEMINI_API_KEY = os.getenv("GEMINI_API_KEY", os.getenv("GOOGLE_API_KEY", ""))
-LLM_API_KEY = os.getenv("LLM_API_KEY", GEMINI_API_KEY or os.getenv("GROQ_API_KEY", ""))
-LLM_BASE_URL = os.getenv("LLM_BASE_URL", "https://generativelanguage.googleapis.com/v1beta/openai/" if (GEMINI_API_KEY or not os.getenv("GROQ_API_KEY")) else "https://api.groq.com/openai/v1")
-LLM_MODEL = os.getenv("LLM_MODEL", "gemini-2.5-flash" if ("googleapis.com" in LLM_BASE_URL or GEMINI_API_KEY or not os.getenv("GROQ_API_KEY")) else "llama-3.3-70b-versatile")
+# Tier-1 Primary: Groq High-Speed Llama-3.3 / Mixtral API (~150ms)
+GROQ_API_KEY = os.getenv("GROQ_API_KEY", os.getenv("LLM_API_KEY", ""))
+LLM_API_KEY = os.getenv("LLM_API_KEY", GROQ_API_KEY)
+LLM_BASE_URL = os.getenv("LLM_BASE_URL", "https://api.groq.com/openai/v1")
+LLM_MODEL = os.getenv("LLM_MODEL", "llama-3.3-70b-versatile")
 LLM_TIMEOUT_SECONDS = 15.0
 
 # Tier-2 & Tier-3 Backup: Cerebras High-Speed LPU (120B model for high instruction following)

llm_synthesis/llm_fallback.py

@@ -79,7 +79,6 @@ class LLMAdapter:
         headers = {
             "Content-Type": "application/json",
             "Authorization": f"Bearer {api_key.strip()}",
-            "x-goog-api-key": api_key.strip(),
             "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) VoiceRAG/1.0",
         }
         
@@ -100,7 +99,7 @@ class LLMAdapter:
         """
         Generate synthesized response grounded strictly in provided context passages.
         Cascade Hierarchy:
-        1. Primary: Gemini Flash / Main LLM (e.g. gemini-2.5-flash, gemini-2.0-flash)
+        1. Primary: Groq Llama-3.3 70B / Mixtral (~150ms)
         2. Tier-2 Backup: Cerebras Gemma 4 31b (gemma-4-31b)
         3. Tier-3 Backup: Cerebras GPT-OSS 120b (gpt-oss-120b)
         4. Tier-4 Local: Deterministic multi-passage extractive fallback
@@ -108,16 +107,14 @@ class LLMAdapter:
         if not config.ALLOW_NETWORK_CALLS_IN_PIPELINE:
             return self._local_fallback_synthesize(prompt, context)
 
-        # Tier 1: Primary LLM (Gemini Flash / OpenAI-compatible)
+        # Tier 1: Primary LLM (Groq Llama-3.3 / OpenAI-compatible)
         if self.api_key and self.api_key.strip():
-            # Try configured model (with instant flash fallbacks on 429)
+            # Try configured Groq model with ultra-fast fallbacks on 429
             primary_models = [self.model]
-            if "googleapis.com" in self.base_url or "gemini" in self.model.lower():
-                for alt_m in ["gemini-2.5-flash", "gemini-2.0-flash", "gemini-1.5-flash"]:
+            if "groq.com" in self.base_url or "llama" in self.model.lower() or "groq" in self.model.lower():
+                for alt_m in ["llama-3.3-70b-versatile", "llama-3.1-8b-instant", "mixtral-8x7b-32768", "llama3-70b-8192"]:
                     if alt_m not in primary_models:
                         primary_models.append(alt_m)
-            elif "groq.com" in self.base_url and self.model != "llama-3.1-8b-instant":
-                primary_models.append("llama-3.1-8b-instant")
                 
             for m in primary_models:
                 res = self._call_chat_endpoint(