Repository navigation
Conversation
list_datasets returned about 73,000 characters for 31 datasets, and the agent reads it on every later model request of a turn (often calling it twice). Caveats shared by several datasets are now stated once, and lineage and column plumbing are left to describe_dataset. The same listing is about 20,000 characters. The MCP tool is unchanged.
One question is several model requests, each re-sending the tool definitions, the instructions, and every earlier tool result. Mark all three as cacheable through OpenRouter so later requests read them at the cached rate. Other providers ignore the markers. The turn's usage now reports cache read and write tokens.
Every question began with a list_datasets call, often two or three. The compact list now sits at the end of the instructions, built once at startup: that saves a model request per question, and the list is part of the prefix Claude reads from cache on every request. If the warehouse can't be read at startup the agent keeps the tool instead.
duaneatat
changed the title
With the list at the end, the rules for charts and answers sat ahead of 20,000 characters of JSON. Three eval repeats scored 86% that way and 89% with the rules last, at the same cost.
Three runs of Claude Haiku 5.5 on the current warehouse: main as a same-day baseline, the list at the end of the instructions, and the list ahead of the rules. List first scores 89% against main's 91% at $0.0039 per chat instead of $0.0146.
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Cuts what the chat model reads and re-reads on every question. Includes the caching change from #183, which this replaces.
list_datasetstool call into the instructions, ahead of the rules, built once at startupusagereportscache_read_tokensandcache_write_tokenslist_datasetstool. The MCP's tool is unchangedEval (Haiku 5.5, 3 repeats, 15 chats, same warehouse; full entry in
docs/reports/chat-eval-history.md):mainthe same day; judge quality 7.4 against 7.5; no errorsmainuses the decennial series. Possible cause, untested: the list describes that series as having "206 missing years"main); the rankings were correct