Summary:
This tweaks the prompting so that the user's message is given on its own instead of as a docstring within Python. This is so that the prompt makes sense when:
- the user asks a question such as "Can you write me a formula which does ...?" rather than describing their formula as a docstring would, or
- the user sends a message that doesn't ask for a formula at all (https://grist.slack.com/archives/C0234CPPXPA/p1687699944315069?thread_ts=1687698078.832209&cid=C0234CPPXPA)
Also added wording for the model to refuse when the user asks for something that the model cannot do.
Because the code (and maybe in some cases the model) for non-ChatGPT models relies on the prompt consisting entirely of Python code produced by the data engine (which no longer contains the user's message) those code paths have been disabled for now. Updating them now seems like undesirable drag, I think it'd be better to revisit this when iteration/experimentation has slowed down and stabilised.
Test Plan:
Added entries to the formula dataset where the response shouldn't contain a formula, indicated by the value `1` for the new column `no_formula`.
This is somewhat successful, as the model does refuse to help in some of the new test cases, but not all. Performance on existing entries also seems a bit worse, but it's hard to distinguish this from random noise. Hopefully this can be remedied in the future with more work, e.g. automatic followup messages containing example inputs and outputs.
Reviewers: paulfitz
Reviewed By: paulfitz
Subscribers: dsagal
Differential Revision: https://phab.getgrist.com/D3936
Summary:
The previous code for extracting a Python formula from the LLM completion involved some shaky string manipulation which this improves on.
Overall the 'test results' from `runCompletion` went from 37/47 to 45/47 for `gpt-3.5-turbo-0613`.
The biggest problem that motivated these changes was that it assumed that code was always inside a markdown code block
(i.e. triple backticks) and so if there was no block there was no code. But the completion often consists of *only* code
with no accompanying explanation or markdown. By parsing the completion in Python instead of JS,
we can easily check if the entire completion is valid Python syntax and accept it if it is.
I also noticed one failure resulting from the completion containing the full function (instead of just the body)
and necessary imports before that function instead of inside. The new parsing moves import inside.
Test Plan: Added a Python unit test
Reviewers: paulfitz
Reviewed By: paulfitz
Subscribers: paulfitz
Differential Revision: https://phab.getgrist.com/D3922
* add support for conversational state to assistance endpoint
This refactors the assistance code somewhat, to allow carrying
along some conversational state. It extends the OpenAI-flavored
assistant to make use of that state to have a conversation.
The front-end is tweaked a little bit to allow for replies that
don't have any code in them (though I didn't get into formatting
such replies nicely).
Currently tested primarily through the runCompletion script,
which has been extended a bit to allow testing simulated
conversations (where an error is pasted in follow-up, or
an expected-vs-actual comparison).
Co-authored-by: George Gevoian <85144792+georgegevoian@users.noreply.github.com>
Summary:
Porting script that run an evaluation against our formula dataset.
To test you need an openai key (see here: https://platform.openai.com/)
or hugging face (it should work as well), then checkout the branch and run
`OPENAI_API_KEY=<my_openai_api_key> node core/test/formula-dataset/runCompletion.js`
Test Plan:
Needs manually testing: so far there is no plan to make it part of CI.
The current score is somewhere around 34 successful prompts over a total of 47.
Reviewers: paulfitz
Reviewed By: paulfitz
Subscribers: jarek
Differential Revision: https://phab.getgrist.com/D3816
