2 ms·
You don't actually need to feed the entire core state as part of each prompt. You just need to tell it that X is one of the things that are known, and if it nee
by int_19h 4y ago
You don't actually need to feed the entire core state as part of each prompt. You just need to tell it that X is one of the things that are known, and if it needs to know X to produce output, it should ask by outputting some F(X) - few-shot learning is usually the most reliable. Then you intercept those in the output and replace them with actual values before feeding it back into LLM for the next prediction.
The "interface" can vary. For example, you can give it a bunch of SQL DDL definitions with descriptive table and field names and tell it that it can query that database as needed. When it works, it can get really impressive - I've seen it do inner queries and joins, and it routinely uses ORDER BY and LIMIT. But then sometimes it just mixes up fields from different tables even in very simple queries (although if you respond with "ERROR: no such field" etc, it can usually self-correct), and it can sometimes interpret the results in creative ways. The approach where it has to query for stuff one by one in a simple well-defined format is much more reliable.