Summary:
Replaces https://phab.getgrist.com/D3940, particularly to avoid doing potentially unwanted things automatically.
Adds optional fields `evaluateCurrentFormula?: boolean; rowId?: number` to `FormulaAssistanceContext` (part of `AssistanceRequest`). When `evaluateCurrentFormula` is `true`, calls a new function `evaluate_formula` in the sandbox which computes the existing formula in the column (regardless of anything the AI may have suggested) and uses that to generate an additional system message which is added before the user's message. In theory this could be used in an interface where users ask why a formula doesn't work, including possibly a formula suggested by the AI. For now, it's only used in `runCompletion_impl.ts` for experimenting.
Also cleaned up a bit, removing `_chatMode` which is always `true` now, and uses of `regenerate` which is always `false`.
Test Plan: Updated `runCompletion_impl` to optionally use the new feature, in which case it now scores 51/68 instead of 49/68.
Reviewers: paulfitz
Reviewed By: paulfitz
Differential Revision: https://phab.getgrist.com/D3970
Summary: Also fixes a few bugs with some telemetry events not being recorded.
Test Plan: Manual.
Reviewers: paulfitz
Reviewed By: paulfitz
Differential Revision: https://phab.getgrist.com/D3960
Summary:
In a nutshell:
- More specific and helpful error messages are shown to the user
- API requests are only retried when needed
- The system deals with reaching the maximum token limit better, especially by switching to a model with a bigger limit
In more detail:
- `COMPLETION_MODEL` configuration has been removed. By default `gpt-3.5-turbo-0613` is used which accepts 4k tokens. If that's not enough, `gpt-3.5-turbo-16k-0613` is used instead.
- Switching to the bigger model happens when either the prompt is too long by itself (the API immediately returns an error code) or the model reaches the 4k limit itself in the process of generating a response and thus returns an incomplete response. The latter case is made possible by removing the `max_tokens: 1500` in the request, which was very generous and would have lead to switching to the more expensive model more often than needed. The downside is that the user has to wait a bit longer for the response.
- If the bigger 16k token limit is also exceeded, the assistant immediately responds (instead of retrying as before) with an error message including suggestions. The suggestions include restarting the conversation if and only if the user has sent multiple messages.
- If a request fails because Grist has reached its OpenAI monthly billing quota, the assistant immediately responds (instead of retrying as before) with an error message suggesting that the user try again tomorrow.
- If a request fails for some other reason, the assistant retries, and if all attempts fail then the user is told to try again in a few minutes and is shown the exact error message, including the API response if there is one.
- Retrying only happens when an API request fails, whereas previously the system also retried errors from a much bigger scope which included calls to the sandbox. The downside is that the hugging face assistant no longer retries, although that code is currently disabled anyway.
- The assistant no longer waits an additional second after the final retry attempt fails.
Test Plan: Added a new server test file with several unit tests using faked OpenAI responses, including the happy path which wasn't really tested before.
Reviewers: dsagal
Reviewed By: dsagal
Subscribers: dsagal
Differential Revision: https://phab.getgrist.com/D3955
Summary:
Adding limits for AI calls and connecting those limits with a Stripe Account.
- New table in homedb called `limits`
- All calls to the AI are not routed through DocApi and measured.
- All products now contain a special key `assistantLimit`, with a default value 0
- Limit is reset every time the subscription has changed its period
- The billing page is updated with two new options that describe the AI plan
- There is a new popup that allows the user to upgrade to a higher plan
- Tiers are read directly from the Stripe product with a volume pricing model
Test Plan: Updated and added
Reviewers: georgegevoian, paulfitz
Reviewed By: georgegevoian
Subscribers: dsagal
Differential Revision: https://phab.getgrist.com/D3907
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
Summary:
Adding a way to detach an editor. Initially only implemented for the formula editor, includes redesign for the AI part.
- Initially, the detached editor is tight with the formula assistant and both are behind GRIST_FORMULA_ASSISTANT flag, but this can be relaxed
later on, as the detached editor can be used on its own.
- Detached editor is only supported in regular fields and on the creator panel. It is not supported yet for conditional styles, due to preview limitations.
- Old code for the assistant was removed completely, as it was only a temporary solution, but the AI conversation part was copied to the new one.
- Prompting was not modified in this diff, it will be included in the follow-up with more test cases.
Test Plan: Added only new tests; existing tests should pass.
Reviewers: JakubSerafin
Reviewed By: JakubSerafin
Differential Revision: https://phab.getgrist.com/D3863
* 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:
The feature is behind a flag GRIST_FORMULA_ASSISTANT (must be "true"). But can be enabled in the
developer console by invoking GRIST_FORMULA_ASSISTANT.set(true).
Keys can be overriden in the document settings page.
Test Plan: For now just a stub test that checks if this feature is disabled by default.
Reviewers: paulfitz
Reviewed By: paulfitz
Subscribers: dsagal
Differential Revision: https://phab.getgrist.com/D3815
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
Summary: This is a backend part for the formula AI.
Test Plan: New tests
Reviewers: paulfitz
Reviewed By: paulfitz
Subscribers: cyprien
Differential Revision: https://phab.getgrist.com/D3786
