Chat Engines

Adding memory on top of retrieve -> synthesize

Every QueryEngine call in Lessons 5-7 was stateless: .query() knew nothing about any call before it, ask the same question twice and you'd get the same retrieval and roughly the same answer both times. index.as_chat_engine() wraps the same Index but adds a conversation history the engine consults on every turn, so a follow-up like "what about for existing employees?" can be understood without repeating "vacation days" or "first 90 days" again.

How condense_question mode works

The default chat mode first asks the LLM to rewrite the newest message into a standalone question using the chat history, turning something like "what about for existing employees?" into something closer to "how much vacation can existing employees use after their first 90 days?" internally, THEN runs that rewritten question through the same retrieve -> synthesize pipeline Lesson 5 used. This is why a ChatEngine can answer a follow-up a plain QueryEngine can't: history is folded into what actually gets *retrieved*, not just handed to the LLM as extra text alongside an unrelated retrieval.

The code, piece by piece

chat_engine = index.as_chat_engine()

Builds a chat engine from the same VectorStoreIndex every prior lesson in this section used.

for message in turns:
    response = chat_engine.chat(message)

The same chat_engine object is reused across every iteration of this loop, on purpose, .chat() mutates the engine's internal history, so reusing one instance is what makes turn 2 aware of turn 1.