A follow-up needs its subject back.
Six three-turn Larkspur Cloud support chats over the twelve handbook chunks from the M10 lab, ranked with the same real TF-IDF. Choose what the retriever receives on turns two and three, then read the hit rate.
- Turn 2 hits
- Turn 3 hits
- Follow-up hit rate
- Rewrite calls to the LLM
The conversation
TF-IDF cosine, top-k shown
A hit means an evidence chunk is in the top k with a score above zero. The customer messages are hand-written; the rewrites are representative of what a rewrite prompt like lesson 33.4's returns, written by hand rather than generated.
All three choices at this k
| Retriever receives | Turn 2 | Turn 3 | Hit rate |
|---|
Twelve follow-ups: turns two and three of six chats. This is the comparison the M33 lab brief asks for, on a small scale.