Release canonical "stable" versions (don't use floating tags)

[macropin]

TLDR, tried installing via Docker and almost everything is broken. Too many bugs to report. Need a stable release that encapsulates all dependencies.

The Docker installation method (r2r serve --docker) uses floating tags for all containers, including 3rd party containers. This means there is no stable release bundle that encapsulates all dependencies, and it's possible that two installs might pull down different container versions, since it's literally the 'latest' container in the registry, not a requested version:

REPOSITORY                                      TAG                 IMAGE ID       CREATED         SIZE
ragtoriches/prod                                3.3.18              71725347deed   5 days ago      1.08GB
ragtoriches/unst-prod                           latest              6293dab7da5d   6 days ago      7.95GB
ragtoriches/cluster-prod                        latest              9c7366466542   6 days ago      2.44GB
emrgntcmplxty/r2r-dashboard                     latest              30cb7acb227a   8 days ago      257MB
ghcr.io/hatchet-dev/hatchet/hatchet-dashboard   latest              11f112b2ac1c   2 weeks ago     131MB
ghcr.io/hatchet-dev/hatchet/hatchet-admin       latest              a549b523f6f8   2 weeks ago     270MB
ghcr.io/hatchet-dev/hatchet/hatchet-engine      latest              caa25a08d232   2 weeks ago     250MB
ghcr.io/hatchet-dev/hatchet/hatchet-migrate     latest              cf47dd659ff4   2 weeks ago     356MB

Please consider updating the compose configurations to use known stable and tested versions of all dependencies.

[NolanTrem]

Hey @macropin the CLI command that you shared should pull the same version Docker container. I agree though, that it is much cleaner to just pin the correct tag and I think we'll look into moving to that direction.

Looking at when these were published, I believe everything here should be stable. Are these bugs preventing the server from starting? Are there certain workflows that you're running where I can replicate some of these bugs? Our CI/CD has been in a passing state, and I haven't had many other issues reported, but would like to get to the bottom of why it seems to be in a bad state for you.

[macropin]

Are these bugs preventing the server from starting?

Yes, the lite install using the default config had issues with some of the containers starting due to (IIRC) configuration issues (possibly missing env vars). The full version did a little better. The r2r dashboard also had a consistent issue with the logs viewer not working, and there were some other inconsistent issues where agent / chat would break for no apparent reason. Graph generation never worked for either.

[NolanTrem]

Can you share some specs around your OS and environment?

[macropin]

It should be nothing unusual. I'm running Jammy Ubuntu 22.04.5 LTS. I installed the python package from PyPi in a virtual env, using the system python (3.10.12). Cuda drivers from nvidia, and ollama manually installed. 128G ram, 2TB NVME, Tesla p40 GPU, AMD Ryzen 7600x CPU.

[macropin]

It would also be helpful to use git tags to track releases. I've noted that r2r==3.3.24 is the latest in PyPi, but there is no corresponding tag or GitHub release for that version.

Also the release does seems broken for me. Running the following:

python3 -m venv venv
./venv/bin/pip install r2r==3.3.24
cp venv/lib/python3.10/site-packages/core/configs/full_local_llm.toml ./
./venv/bin/r2r serve --docker --full --config-path=/home/andrew/code/r2r/full_local_llm.toml

Gives the following result

Traceback (most recent call last):
  File "/home/andrew/code/r2r/./venv/bin/r2r", line 5, in <module>
    from cli.main import main
  File "/home/andrew/code/r2r/venv/lib/python3.10/site-packages/cli/main.py", line 7, in <module>
    from cli.command_group import cli
  File "/home/andrew/code/r2r/venv/lib/python3.10/site-packages/cli/command_group.py", line 6, in <module>
    from typing import Any, Never
ImportError: cannot import name 'Never' from 'typing' (/usr/lib/python3.10/typing.py)

[NolanTrem]

Thanks @macropin, sorry for not following up here sooner.

We'll be moving to canonical versions on images soon. This is also a great catch—we were just about to cut a release, so I've gone ahead and fixed the import of Never to support Python 3.10. Will also generate a GitHub release.

There should be some bug fixes here that might be related to the streaming issues that you saw. I've also made some improvements to the graph generation process, that may have been the root of your troubles if you were using local models. This might be of interest: https://r2r-docs.sciphi.ai/cookbooks/local-llms#extracting-entities-and-relationships

Overall, from what we've seen and heard from users, the system should be in a good state. Please let me know if you have any issues with the latest.

[macropin]

Overall, from what we've seen and heard from users, the system should be in a good state. Please let me know if you have any issues with the latest.

Thanks for your help here. It's mostly working now with 3.3.27, and I've been able to kick the tyres and see what it's all about!

A few issues I've spotted so far:

• The r2r cli isn't working anymore for me. ./venv/bin/r2r documents create .... fails with "Request failed: [Errno -5] No address associated with hostname" which looks like a configuration issue when running in docker, but I'm not sure where to look to fix this.
• In the UI the RAG agent returns Json responses, that are more or less an echo of the question asked. Not sure what to expect?
• Some stability issues, seemingly caused by ollama, but that is not your problem.