In the current CLI implementation, default values for the model,
temperature, and history directory were provided as defaults to the
`typer` CLI interface. This bypassed the ability to set values via
environment variables and relied on `typer` rather than `pydantic` to
validate the history directory. This prevented the directory's creation
when it doesn't exist.
This commit sets the default values in the CLI to `None` and thus relies
on the `OpenAISettings` class for default values and parameter
validation.
When updating the model pricing in commit 50fa0cc5ae I forgot to update
the associated test. As well as fixing this test, this commit also
updates the calculation to use `math.floor` instead of `round` to round
to six decimal places. This is because the `round` function appeared to
round incorrectly. For example, when running the test, 0.0000275 was
rounded to 0.000028 instead of the expected 0.000028.
The ChatFake object used in CLI tests never gets a `conversation`
attribute defined. This resulted in multiple tests failing since the
`save_conversation` function accessed the `name` attribute via
`current_chat.conversation.name`. This was resolved by adding a `name`
property to the `Chat` class, which can be easily faked in tests.
Additionally, formatting was fixed.
This commit adds the additional CLI command `load`, which accepts the
path to a previously saved conversation from which to load previous
messages. The previous cost and token counts are not loaded as theses
are deemed functions of the chat session rathr than the conversation
and are stored with conversation purely for future reference.
At the end of each sesssion the conversation is stored to a directory
(defaulting to `.history` in the currrent working directory) as a JSON
object. Note, the session must be ended by sending the quit message (/q)
for the conversation to be saved. Ctrl+C will not work.
This commit enables the user to provide one or more text files as
context for their chat session. These will be provided as system
messages to OpenAI's API, one message per file.
Implements a basic CLI using Typer, Rich, and Prompt Toolkit. I couldn't
work out how to propoerly mock/fake the Rich console or the prompt
session from Prompt Toolkit, so I created dummy functions to work around
them and just test the basic application flow rather than user inputs
and outputs to the terminal.
Linting checks are performed using Ruff, Mypy, and Pydocstyle. Code
formatting is performed using Black and Isort. A Makefile is also added
with linting and formatting commands for convenience.
This commit implements a basic interface to hold a conversation with
OpenAI's GPT models via their API, using OpenAI's Python SDK. In
addition to this interface, the commit incorporates Pydantic for
managing settings via environment variables.
Paul Harrison