GitHub user eugenegujing added a comment to the discussion: Proposal: Making 
the workspace agent useful to a new user by seeing run errors, reading the 
docs, knowing the datasets

Thanks everyone for the feedback. Here is where things stand.

The PRs are open:

- #9423 reports the user's run to the agent (change 1)
- #9424 adds the docs search and read tools (change 2)
- #9426 opens links in agent chat messages in a new tab. It is small and 
separate, and it matters once the agent answers with docs links.

@chenlica I added three diagrams to the post above: an overview, and one flow 
for each of the first two changes. The PRs carry more detailed ones.

@zyratlo Both of your points are in. #9423 sends each operator's state and its 
input and output row counts from the user's last run, along with the errors, so 
the agent can see that a file scan produced 0 rows. On docs, you were right 
that the tools only help as far as the pages go. One example from testing: 
`tutorials/create-dataset-upload-data.md` still tells the user to open an 
environment tab and click Add Dataset, which the editor no longer has, and the 
agent repeated that step. The search evaluation in #9424 uses 25 questions I 
wrote myself, so your collaborator's questions might also be a better test set.

@aicam I agree with unifying the endpoints the user and the agent use, and I 
will post the agent's requirements in #9395. I don't think the open PRs 
conflict with that effort. Neither of them changes backend code or adds a call 
to a backend endpoint:

- #9423 changes only the frontend and the agent service. It takes the run's 
errors and row counts from the browser, because nothing serves them to the 
agent service today. The agent service only receives a validated report, so 
once #9395 lets it read a run the way the user does, the source can be switched 
without changing how the agent uses the report.
- #9424 reads the repository's `docs/` folder.
- Change 3 (dataset tools) is the one that depends on #9395. It is not started, 
and I will wait for that design instead of building it on the current 
file-service endpoints.

GitHub link: 
https://github.com/apache/texera/discussions/9377#discussioncomment-19604694

----
This is an automatically sent email for [email protected].
To unsubscribe, please send an email to: [email protected]

Reply via email to