examples: document how to log chats

[nbsp]

mostly copying the minimal assistant example, with this section added:

async def log_context():
    with open("transcriptions.log", "w") as f:
        chat_ctx = assistant.chat_ctx.copy()
        for msg in chat_ctx.messages:
            f.write(msg.role + "\n")
            f.write(msg.content + "\n\n")

ctx.add_shutdown_callback(log_context)

[changeset-bot]

⚠️ No Changeset found

Latest commit: 5a5d52f

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

This PR includes no changesets

When changesets are added to this PR, you'll see the packages that this PR includes changesets for and the associated semver types

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

[theomonnom]

let's avoid blocking ops.
we could also just get the assistant.chat_ctx in a shutdown callback and write everything at once?

[davidzhao]

hmm I think we should not assume that everything will be buffered. it's better to demonstrate control where user may choose how to buffer

[nbsp]

so should we have both a shutdown log and a real-time log?

[davidzhao]

for example, what if I'd like to send transcriptions to my backend service on the fly to log? (in order to improve latency and minimize risk of losing data).

for really long sessions (think 2h+) we also don't want to buffer everything in memory.

ideally we'd also expose timing data (by emitting them in realtime, we can basically avoid having to tag timestamps)

additionally, some users may use summarization techniques with the LLM, so the chatcontext isn't an accurate representation of each sentence being said by the user/agent.

[nbsp]

new commit reverts the changes, going back to realtime, and also makes it non-blocking.
i also print the current datetime in the header for each message.

[davidzhao]

what about using a queue here instead of opening the file each time?

IMO this would be a cleaner design:

• log_user_speech and log_agent_speech push a dict with details to a asyncio.Queue
• spin up only a single task to write to file
• the task should flush periodically and close itself when disconnected
• we can avoid creating separate tasks for these callbacks due to synchronization issues (i.e. if both things write at the same time, chatlog is corrupted)

[nbsp]

sure, but this isn't production code. it shows that the agent and user texts are reachable like this and then you can do whatever with them: log to a file, append to a database, forward to somewhere else, etc. minimal_assistant isn't a demo you're gonna run on prod, having a lot of syscalls doesn't matter here imo

[nbsp]

this now pushes to a queue, which is being read from continuously in a non-blocking way. when the room closes so does the queue.

[theomonnom]

task_done is doing nothing on its own. This is only relevant when using queue.join().

I would also prefer to use ctx.add_shutdown_callback instead of room disconnect

[davidzhao]

lg

[theomonnom]

nit: rename the file to save_transcriptions? or save_chatctx?