Can you elaborate on a project where you integrated LLMs using RAG (Retriever‑Augmented Generation)? What challenges did you face during integration and how did you overcome them?

In a recent project I built a knowledge‑base chatbot for a customer support portal. The core was a RAG pipeline: first, we indexed a corpus of internal documents using OpenAI embeddings stored in Pinecone. The retrieval step used cosine similarity to fetch the top‑k relevant passages. The generator was a fine‑tuned GPT‑3.5 model that received the query and the retrieved snippets as context. Challenges included latency – the retrieval step added ~200 ms, and the generator added ~500 ms, pushing the total response time over our 1 s SLA. We mitigated this by caching frequent queries, using a smaller embedding model (Ada) for retrieval, and batching requests. Another issue was hallucination; the model sometimes generated facts not present in the retrieved docs. We added a post‑generation verifier that cross‑checked key facts against the source snippets. Finally, we had to integrate the pipeline into an existing Spark ETL that updated the vector store nightly. We used Airflow to orchestrate the nightly refresh and a lightweight Flask API to expose the RAG service. The result was a robust, low‑latency chatbot that improved first‑contact resolution by 15 %.