""" RAG query engine with hybrid retrieval + reranking (LlamaIndex). Pipeline: question -> retriever (wide candidate pool) -> reranker (narrows to best matches) -> query engine -> Claude -> answer Install (add to requirements.txt): llama-index-core>=0.11.0 llama-index-llms-anthropic>=0.3.0 llama-index-embeddings-huggingface>=0.3.0 sentence-transformers>=3.0.0 """ import os from llama_index.core import Settings, VectorStoreIndex, SimpleDirectoryReader from llama_index.core.postprocessor import SentenceTransformerRerank from llama_index.core.query_engine import RetrieverQueryEngine from llama_index.embeddings.huggingface import HuggingFaceEmbedding from llama_index.llms.anthropic import Anthropic # --------------------------------------------------------------------------- # 1. GLOBAL SETTINGS — LLM + embedding model, set once for the whole app. # --------------------------------------------------------------------------- # Reads ANTHROPIC_API_KEY from environment — set this as a secret in your # hosting environment, do not hardcode it. Settings.llm = Anthropic( model="claude-sonnet-4-6", # match whichever Claude model you're using today api_key=os.environ["ANTHROPIC_API_KEY"], ) # Swap model_name to match whatever sentence-transformers model you're # already running for embeddings. Settings.embed_model = HuggingFaceEmbedding(model_name="BAAI/bge-base-en-v1.5") # --------------------------------------------------------------------------- # 2. INDEX — build once from your document store, cache the result. # --------------------------------------------------------------------------- def build_index(data_dir: str = "data") -> VectorStoreIndex: """ Loads your content and builds a LlamaIndex vector index over it. If you already have embeddings stored somewhere (e.g. an existing vector store), replace this with the matching LlamaIndex vector store loader instead of rebuilding from raw files. """ documents = SimpleDirectoryReader(data_dir).load_data() index = VectorStoreIndex.from_documents(documents) return index # --------------------------------------------------------------------------- # 3. RERANKER — narrows a wide candidate pool down to the most relevant few. # --------------------------------------------------------------------------- # Small cross-encoder: runs on CPU, no GPU required, fast enough for # real-time use on a handful of candidates. top_n is how many survive # reranking and get passed to Claude. reranker = SentenceTransformerRerank( model="cross-encoder/ms-marco-MiniLM-L-6-v2", top_n=5, ) # --------------------------------------------------------------------------- # 4. QUERY ENGINE — retrieval (wide) + reranking (narrow) + Claude, wired together. # --------------------------------------------------------------------------- def build_query_engine(index: VectorStoreIndex) -> RetrieverQueryEngine: """ similarity_top_k=20 -> retrieve a wide candidate pool (this is the pool the reranker gets to work with; too narrow here defeats the point of reranking). node_postprocessors=[reranker] -> automatically reranks those 20 down to the top 5 before building the prompt sent to Claude. """ retriever = index.as_retriever(similarity_top_k=20) query_engine = RetrieverQueryEngine.from_args( retriever=retriever, node_postprocessors=[reranker], ) return query_engine # --------------------------------------------------------------------------- # STREAMLIT INTEGRATION EXAMPLE # --------------------------------------------------------------------------- # import streamlit as st # from rag_engine import build_index, build_query_engine # # @st.cache_resource # def get_query_engine(): # index = build_index("data") # your content folder # return build_query_engine(index) # # query_engine = get_query_engine() # # question = st.text_input("Ask a question") # if question: # response = query_engine.query(question) # st.write(str(response)) # # # Optional: show which sources were used, post-rerank # with st.expander("Sources"): # for node in response.source_nodes: # st.write(f"- {node.node.metadata.get('file_name', 'unknown')} " # f"(score: {node.score:.3f})")