task(BE-5766): Introduce PAPI-HETA tools#412
Conversation
|
Plan Error |
b718c11 to
34e633f
Compare
221f3e1 to
ab631d9
Compare
| Client._api_client_uncached = None | ||
|
|
||
|
|
||
| def _retry_on_auth_failure(func: Callable[P, R]) -> Callable[P, R]: |
There was a problem hiding this comment.
is it possible for this function to retry indefinitely if something goes really wrong and auth keeps failing?
There was a problem hiding this comment.
Hmm, I don't think so, at least not due to the decorator implementation. The decorator only clears the cached token and then retries the wrapped function once.
There could be a hang if something goes wrong in authentication when a Client() is instantiated but that wouldn't be due to this retry.
src/aignostics/mcp/_server.py
Outdated
|
|
||
| lines = [header, separator] | ||
| for i, row in enumerate(rows): | ||
| if i >= MAX_SQL_RESULT_ROWS: |
There was a problem hiding this comment.
wouldn't it make more sense to add LIMIT to the sql query?
There was a problem hiding this comment.
Hmm, what I wanted to do is have a way to return limited number of rows in the tool result but also tell the LLM that there were more rows in the output of the SQL query. So that the LLM can "know" that there were more rows which couldn't be returned (say if the query was returning 200k rows) and that it needs to take a different approach.
But there's a better way to do this. I'll change this with some other changes I'll push in a short while about the handling of tool results. I realized we need some safeguards in the output of the visualize calls.
Thanks 🙂
There was a problem hiding this comment.
Changes have been pushed. Needed to do some more around the outputs of tool calls.
ddb2128 to
3dc9e69
Compare
d57c928
into
task/BE-5757-central-mcp-server-python-sdk
No description provided.