Update env-vars.md with new variables for local hosted models

[wirthual]

Added OpenAI variables which are required to run the example with locally hosted models.

Example got tested with locally hosted embedding model:

docker run -it -p 7997:7997  michaelf34/infinity:latest  v2  --model-id mixedbread-ai/mxbai-embed-large-v1 --device cpu --engine optimum

Together with a chat completion model:

ollama run llama3.2:1b

I was able to execute the Getting Started guide with the following command:

OPENAI_BASE_URL='http://localhost:7997' OPENAI_ENDPOINT='http://localhost:11434/v1/chat/completions' OPENAI_API_KEY='does-not-matter' OPENAI_MODEL="llama3.2:1b" uv run python code.py

Which resulted in the following output:

0.044s -- Using OpenAI
Indexing 3 messages...
Indexed 3 messages.
Got 15 semantic refs.

The used model uses embedding dimensions of 1024, which I adapted by changing DEFAULT_EMBEDDING_SIZE

[gvanrossum]

Thanks! I have some suggestions. After applying and pushing those I will merge your PR.

[gvanrossum]

Can I get this blank line back? I like the extra bit of whitespace in the rendered page.

[wirthual]

Done.

[gvanrossum-ms]

Thanks for your contribution! Really cool you got it to work with a local model.

Regarding the embedding width, I think the correct way currently is to update the dict model_to_embedding_size_and_envvar in typeagent/aitools/embeddings.py with an appropriate entry (you could patch it in from the outside, that variable is not private) and then use code like this:

from typeagent.aitools.embeddings import AsyncEmbeddingModel
from typeagent.knowpro.convsettings import ConversationSettings

async def main():
    embedding_model = AsyncEmbeddingModel(embedding_size=1024, model_name="your-name")
    settings = ConversationSettings(model=embedding_model)
    settings.semantic_ref_index_settings.auto_extract_knowledge = True      # DON'T FORGET THIS
    conversation = await create_conversation("demo.db", TranscriptMessage, settings=settings)
    # ... original code ...

You could argue that it should be possible to construct an AsyncEmbeddingModel instance that doesn't correspond to a table entry. It looks like there's already an argument endpoint_var but no logic that uses it. If you're interested I encourage you to submit a PR.