User should be able to toggle additional resources for context in Inline completions settings

ryderwishart commented 3 months ago

Refs #3

[x] We need to pull in the source text into the inline suggestions as few-shot pairing.
(edit: we are not diong this) We should generate two suggestions each time: 1 for completing the line only (stop at VREF or newline), 2 for continuing as long as possible.
[x] Few-shot prompt
[ ] Master prompt (edit: is this useful?)

thing to include: 

[x] pull in parallel passages as verse context pairs
[ ] (might need more thought) pull in Verses with similar overlap in dictionary words or dictionary list / a gloss (not created yet)
[x] (might need more thought) Option to include Other parallel texts that are not the bible (this could be a simple json structure in a defined file of source: "", target: "" pairs that the llm can optionally pull in as context)

DoD

Users should be able to turn on additional Bible context for the inline completions. This will result in the user having better suggestions.

Details:

This will give the model some few shot examples in order to produce better suggestions based on the project context.

BenjaminScholtens commented 3 months ago

We should pull in words/tokens from the dictionary that are likely to be included in the verse being translated. AKA. If GEN 1:1 and GEN 3:3 have a lot of the same words in them we can pull the likely glosses from the dictionary and add them to the prompt. e.x. "Past translation chunks that had similar source words used these target words in the translation [..., ..., ..., ...] Here are the words in context with the verse pairs, [ {source: ..., target: ....}, ... ]. try to translate 'x' into the target using the above word suggestions and infer grammar and wording via the translation pairs provided"

BenjaminScholtens commented 3 months ago

Create multiple inline suggestions with different prompts. Collect telemetry about the most used/useful prompt.

ryderwishart commented 2 months ago

One additional feature requested by a user is:

Is there a way to tell the model what books to use and which books not to use?

I think we need

[x] [edit: Daniel and I decided to use a string] a json_database search param for specific book filters
[x] an optional argument in the message passing for the web socket
[x] a configuration in the frontend workspace settings

Anything I'm forgetting?

genesis-ai-dev / codex-comments-and-chat

User should be able to toggle additional resources for context in Inline completions settings #4

DoD

Details: