Problem
A chat assistant that treats every conversation as the first one wastes the user's effort. The user explains something in detail, and next time the assistant asks the same questions again, so the user stops sharing. Yandex's Alice voice assistant has a large "just chatting" scenario on any topic, so a fixed topic list does not fit. The memory must also work on every utterance, including information shared moments ago.
Example from the article: a user says the cat is named Apelsin and is ginger. Later the cat knocks a vase off the table, and the assistant asks "what is his name?" and "what color is he?" instead of replying that ginger cats can be restless.
Solution in brief
Retrieve the parts of past dialogues most relevant to the current context with a trained embedder. Then fine-tune an LLM to summarize them into short self-contained facts, and generate the reply with those facts. Edge cases (nothing new in relevant parts, no context) are handled by training labels and a small prepared fact set.
Questions this case answers
- How to make an LLM assistant's replies personal by retrieving the relevant parts of a user's past conversations and summarizing them into self-contained facts, without a fixed topic list
- How to evaluate personalization of open-ended dialogue replies when quality is subjective
- How to personalize an LLM assistant's replies using facts remembered from past conversations
- How to retrieve relevant parts of a user's dialogue history for the current conversation
Tools and methods
- Embedder over dialogue parts — retrieves the top-K relevant fragments from history
- LLM summarizer — turns retrieved fragments into facts under a list of criteria
- LLM scoring — rates summarized facts against a reference set for precision, recall and F1
- Side-by-side comparison — end-to-end check of personalization against the plain reply
Google or a one-time email code. No password.
