Envoy has proven itself in the industry and we didn't want to reinvent the wheel by doing what envoy had already done for observability, rate-limits, connection management etc. And reason for using proxy-wasm was so we don't take hard dependency on any built version of envoy. There are many other benefits too which are listed here [1].
Regarding support for wasm runtime in envoy. We believe wasm support in envoy is not going anywhere and it will continue to become more and more stable over time. Envoy has heathy community and in case of any security vulnerability we will hope that envoy will ship fix quite fast which it has done in the past. See here for details of security patch rollout [2]
For complex agent scenarios where there might be COT reasoning needed how does archgw work in that scenario? BTW nice detailed post. And congratulations on the launch.
If you are building a state machine then you should use tools that enable you to do that (temporal, langgraph, etc) by orchestrating multple LLM calls - in that case archgw offers intelligent routing to your COT agent and enables you to transparently add rich observability and metrics for all calls made to LLMs in the COT/state machine scenario.
I am interested in knowing how the arch would run? Is it a library that I need to add in my code to make it work? Or do I need to deploy some sort of service in my infrastructure?
Its a proxy - built on Envoy. I think its fairly clear that this is a separate process. As far as I can tell, you create a config file, boot up archgw, and in the config have it point to endpoints where prompts get forwarded.
Thanks! Those are all good questions. Let me respond to them one by one,
> Can I just use arch for routing between LLMs
Yes, you can use arch_config.yaml file to select between LLMs. In fact we have a demo on llm_routing [1] that you can try. Here how you can specify different LLMs in our config,
We currently support mistral and openai. And for both of them we support streaming interface. We do expose openai complaint v1/chat interface so any chat UI that works with openai should work with us as well. We do ship demos with gradio sample application.
> And what about key management? Do I manage access keys myself?
None of your clients need to manage access keys. Upon receipt of request our filter will appropriate LLM from arch_config and pick relevant access_key and modify request with access_key from arch_config before sending request to upstream LLM [2].
This is honestly quite a detailed and thoughtfully put together post. I do have some questions and would love to hear your thoughts on those. First off, can I use just the model itself? Do you have models hosted somewhere or they run locally? If they run locally what are the system requirements? Can I build RAG based applications on arch? And how do you do intent detection in multi-turn dialogue? How does parameter gathering work, is the model capable of conversing with the user to gather parameters?
Arch-Function our fast, open source LLM does most of the heavy lifting on extracting parameter values from a user prompt, gathering more information from the user, determining the right set of functions to call downstream. Its designed for smarter RAG scenarios and agentic workflows (like buying an insurance claim through prompts). While you can use the model yourself, archgw offers a framework on usage to detect hallucinations and re-prompt the LLM if token logprobs.
The same model is currently being updated to handle complex multi-turn intent and parameter extraction scenarios, so that the dreaded follow-up, clarifying RAG use case can be effortlessly handled by developers without having to resort to complex LLM pre-processing. In essence, if the user's follow-up question is "remove X", their RAG endpoints get structured information about the prompt and refined parameters against which developers simply have to retrieve the right chunks for summarization.
Thanks :) thanks for those comments. Those are great questions. Let me respond to them one by one,
> Can I use just the model itself?
yes - our models are on huggingface. You can use them directly.
> Do you have models hosted somewhere or they run locally? If they run locally what are the system requirements?
arch gateway does bunch of processing locally for example for intent detection and hallucination we use nli model. For function calling we use hosted version of our 1.5B function calling model [1]. We use vllm to host our model, But vllm is not supported on mac. There are other issues too running model locally on mac. For example docker doesn't support giving gpu access on mac to containers. We tried using ollama in the past to host model but ollama doesn't support exposing logprobs. But we do have an issue on this [2] and we will improve it soon.
> Can I build RAG based applications on arch?
Yes you can. You would need to host vector db. In arch we don't host vector db, we wanted to keep our infra simple and clean. Do do have a default target that you can use to build RAG application. See this demo for example see insurance agent demo [3]. We do have an open issue on building a full RAG demo here [4], +1 to it to show your support.
> How does parameter gathering work, is the model capable of conversing with the user to gather parameters?
Our model is trained to engage in dialogue if a parameter is missing because our model has seen examples of missing parameters during training. During our evals and tests we found out that our model could still hallucinate e.g. for the question "how is the weather" model could hallucinate city as "LA" even though LA was not specified in query. We handle hallucination detection in arch using nli model to establish entailment of parameters from input query. BTW we are currently working on to improve that part by quite a lot. More on that in next release.
With all the focus on language specific frameworks - this out of process architecture choice is an interesting one. On one hand, it helps you side step the "is this functionality available on js, java, etc" question, and on the other it means its not as easy as `import archgw` in python. Good luck though, feels like an interesting project
… This extension is functional but has not had substantial production burn time, use only with this caveat.
This extension has an unknown security posture and should only be used in deployments where both the downstream and upstream are trusted.
Regarding support for wasm runtime in envoy. We believe wasm support in envoy is not going anywhere and it will continue to become more and more stable over time. Envoy has heathy community and in case of any security vulnerability we will hope that envoy will ship fix quite fast which it has done in the past. See here for details of security patch rollout [2]
[1] https://github.com/proxy-wasm/spec/blob/main/docs/WebAssembl...
[2] https://github.com/envoyproxy/envoy/blob/main/SECURITY.md
Can I just use arch for routing between LLMs? And what LLMs do you support? And what about key management? Do I manage access keys myself?
> Can I just use arch for routing between LLMs
Yes, you can use arch_config.yaml file to select between LLMs. In fact we have a demo on llm_routing [1] that you can try. Here how you can specify different LLMs in our config,
> And what LLMs do you supportWe currently support mistral and openai. And for both of them we support streaming interface. We do expose openai complaint v1/chat interface so any chat UI that works with openai should work with us as well. We do ship demos with gradio sample application.
> And what about key management? Do I manage access keys myself?
None of your clients need to manage access keys. Upon receipt of request our filter will appropriate LLM from arch_config and pick relevant access_key and modify request with access_key from arch_config before sending request to upstream LLM [2].
[1] https://github.com/katanemo/archgw/tree/main/demos/llm_routi...
[2] https://github.com/katanemo/archgw/blob/main/crates/llm_gate...
Do you have to use envoyproxy to use archgw? Can archgw be used for LLM routing without using envoyproxy?
> do you have to use envoyproxy to use archgw
Yes, 100%. Our gateway is implemented as rust filter which runs inside envoy process.
> Can archgw be used for LLM routing without using envoyproxy?
Unfortunately no. Since we are built in top of envoyproxy.
The same model is currently being updated to handle complex multi-turn intent and parameter extraction scenarios, so that the dreaded follow-up, clarifying RAG use case can be effortlessly handled by developers without having to resort to complex LLM pre-processing. In essence, if the user's follow-up question is "remove X", their RAG endpoints get structured information about the prompt and refined parameters against which developers simply have to retrieve the right chunks for summarization.
> Can I use just the model itself?
yes - our models are on huggingface. You can use them directly.
> Do you have models hosted somewhere or they run locally? If they run locally what are the system requirements?
arch gateway does bunch of processing locally for example for intent detection and hallucination we use nli model. For function calling we use hosted version of our 1.5B function calling model [1]. We use vllm to host our model, But vllm is not supported on mac. There are other issues too running model locally on mac. For example docker doesn't support giving gpu access on mac to containers. We tried using ollama in the past to host model but ollama doesn't support exposing logprobs. But we do have an issue on this [2] and we will improve it soon.
> Can I build RAG based applications on arch?
Yes you can. You would need to host vector db. In arch we don't host vector db, we wanted to keep our infra simple and clean. Do do have a default target that you can use to build RAG application. See this demo for example see insurance agent demo [3]. We do have an open issue on building a full RAG demo here [4], +1 to it to show your support.
> How does parameter gathering work, is the model capable of conversing with the user to gather parameters?
Our model is trained to engage in dialogue if a parameter is missing because our model has seen examples of missing parameters during training. During our evals and tests we found out that our model could still hallucinate e.g. for the question "how is the weather" model could hallucinate city as "LA" even though LA was not specified in query. We handle hallucination detection in arch using nli model to establish entailment of parameters from input query. BTW we are currently working on to improve that part by quite a lot. More on that in next release.
[1] https://huggingface.co/katanemo/Arch-Function-1.5B.gguf
[2] https://github.com/katanemo/archgw/issues/286
[3] https://github.com/katanemo/archgw/blob/main/demos/insurance...
[4] https://github.com/katanemo/archgw/issues/287