Add full RAG pipeline: agent, rag_engine, generator, knowledge_base, full Gradio UI

agent.py

@@ -95,14 +95,14 @@ class MLOpsRAGAgent:
             logger.error(f"Retrieval failed: {e}")
             return [], []
 
-        # ChromaDB returns cosine *distance* (lower = more similar).
-        # A score threshold on distance would silently discard the best chunks,
-        # so we pass all retrieved nodes to the Flan-T5 relevance check instead.
         scores = [self.rag.get_node_score(n) for n in nodes]
         logger.info(f"Raw node distances: {[round(s, 3) for s in scores]}")
 
-        relevant = []
-        for node in nodes:
+        # Always keep the top 3 nodes (retriever returns them sorted by distance
+        # ascending, so these are the closest matches).  Run the LLM relevance
+        # check only on nodes 4-6 to optionally widen the context.
+        relevant = list(nodes[:3])
+        for node in nodes[3:]:
             text = self.rag.get_node_text(node)
             try:
                 if self.gen.check_relevance(query, text):