.png)
.gif)
Last updated: March 31, 2023
In recent times, Large Language Models (LLMs) have taken the spotlight in the world of artificial intelligence. However, there seems to be a lack of discussion on how to effectively deploy them in real-world settings. To address this gap, we embarked on a journey to incorporate an LLM into a production-ready workflow. And so, 🥁 …
The ZenML Slack bot is live! 🥳
We couldn’t ignore the growing enthusiasm surrounding LLMs, so we decided to explore how these emerging technologies could seamlessly fit into ZenML’s MLOps workflows and standards. Our solution? We created and deployed a Slack bot to provide community support. Feel free to @zenml-bot in our #general Slack channel to receive assistance with your ZenML-related queries.
First, some context. Large language models (LLMs) have become a cornerstone of natural language processing, offering unparalleled capabilities for knowledge generation and reasoning. The past few weeks have seen a number of high profile releases of models and interfaces. However, despite their immense potential, incorporating custom, private data into these models remains a challenge. This is where tools like LangChain and LlamaIndex (formerly ‘GPT Index’) come into play, offering innovative solutions for data ingestion and indexing, enabling developers to augment LLMs with their unique datasets.
LangChain and LlamaIndex facilitate in-context learning, an emerging paradigm that allows developers to insert context into input prompts, leveraging LLM’s reasoning capabilities for generating more relevant and accurate responses. This differs from finetuning, which requires retraining models using custom datasets, often demanding significant computational resources and time.
By addressing data ingestion and indexing, LangChain and LlamaIndex provide a streamlined framework for integrating custom data into LLMs. Their flexible design simplifies incorporating external data sources, enabling developers to focus on creating powerful applications that harness LLMs’ full potential.
These tools bridge the gap between external data and LLMs, ensuring seamless integration while maintaining performance. By utilizing LangChain and LlamaIndex, developers can unlock LLMs’ true potential and build cutting-edge applications tailored to specific use cases and datasets.
🛣️ The project we built uses both langchain and llama_index as well as some extra code for the Slack bot itself. The rest of this blogpost covers the technical details of what we implemented. If you want to get your hands dirty and try out a simpler version, feel free to check out our Generative Chat example that was released last week.
ZenML 🤝 LLM frameworks
There are various terms being tried out to describe this new paradigm — from LLMOps to Big Model Ops. Not only the words used to describe how engineering will work are new, but the underlying structures and frameworks are also being developed from the ground up. We wanted to witness these changes first hand by participating and getting our hands dirty.
In particular, we wanted to experience how users of ZenML might go about using our framework to integrate with these tools and models. We had seen lots of demos showcasing useful and striking use cases, but none were some of complexities around deploying these in production were raised.
Starting out, we wanted to get a feel for the two well-known options for developing on top of large language models — that is, LangChain and Llama Index — by trying out some of their core functionality. In particular, loading a series of documents, doing some kind of querying of those documents and packaging this all up into some kind of frontend user experience seemed like good places to start.
We approached the problem by dividing it into two main parts, each addressing a specific aspect of the project. The first part focused on creating the ZenML pipeline, while the second part concentrated on the larger workflow that would automate the bot updates.
Part 1: Creating the Index with a ZenML Pipeline
In the first part of the project, we built a ZenML pipeline to handle data ingestion, construct the vector store, and save the resulting data in the ZenML artifact store. The pipeline consisted of the following big-picture actions:
- Data Ingestion: We used the langchain and llama_index frameworks to load a series of documents from various sources such as GitHub and Slack. This data would serve as the basis for our vector store.
- Vector Store Construction: Once the data was ingested, we applied the necessary preprocessing and transformed it into a suitable format for querying. We then used the large language models to create a vector store, which would enable us to search and retrieve relevant information efficiently.
- Saving to ZenML Artifact Store: The final piece was to save the resulting vector store in the ZenML artifact store. This allowed us to have a versioned and easily accessible storage system for our indexed data.
We thought about the kinds of data sources we’d want to be queried when users entered their queries. We take a lot of pride and effort to maintain our documentation so that was an obvious one to include. We considered adding the README files of our examples since that’s often a starting point for our users, and we thought our release notes would also be useful to be part of the context.
Getting all of this together was not too hard, and LangChain and Llama Index released new versions that helped with obtaining Slack messages and documentation from Gitbook. In the end we wanted more flexibility than was available from the pre-built document loaders, so we just used the generic loader for web content along with a custom scraper to get a list of all URLs available for our docs and examples READMEs.
We settled on using a vector store as the way to query our documentation. This seems to be a common way to do this when your texts are quite large. You don’t want to have to send your entire corpus over to OpenAI (or whoever is providing your LLM) since most of it will be irrelevant to the specific question at hand. Instead, what you want is a way to get only the top handful of useful chunks of information and pass them on to your LLM along with a prompt and the original question.
We used the FAISS Vector Store. FAISS (Facebook AI Similarity Search) is a library developed by Facebook AI Research for efficient similarity search and clustering of high-dimensional vectors. It is particularly useful for tasks that involve large-scale information retrieval or finding nearest neighbors in high-dimensional spaces. Moreover, it’s fast and open-source and was quite well documented for this use case. (LangChain and Llama Index have quite a few other options, particularly if you want your vector store hosted and deployed somewhere in the cloud.)
The documents downloaded in earlier steps are split up into small 1000-character chunks, then combined together with embeddings and fed into the vector store. This special combination is what allows us to query a large corpus of text.
Part 2: Deploying the ZenML Bot with an automated workflow
The second part of the project focused on setting up a continuous integration and continuous deployment (CI/CD) workflow to ensure that the bot would be updated automatically whenever a new release was made. This process involved the following steps:
- Triggering the Pipeline: We configured the pipeline to run automatically whenever a new release was made, ensuring that the most up-to-date data would be used for the bot’s functionality.
- Running the ZenML Pipeline: As part of the workflow, the ZenML pipeline would be executed, updating the vector store with the latest data and ensuring that the bot would be able to retrieve the most relevant information for user queries.
- Deploying the Updated Bot: Once the pipeline had successfully run and the updated vector store was saved in the ZenML artifact store, the new version of the bot was automatically deployed to Slack. This ensured that users would always interact with the most current version of the bot, providing the best possible user experience.
The Slack bot itself wasn’t too hard to put together. Our pattern of use was simple and well-defined. The only unusual part was using the vector store that had been generated in our pipeline as the basis for incoming Slack messages. We deployed our bot inside a container along with a way to query ZenML for the vector store on initialization. It’s now up and running courtesy of Google Cloud Run.
In summary, our project combined the power of ZenML with the capabilities of langchain and llama_index to create an automated system that not only indexed data from our documentation but also updated the bot continuously. This approach allowed us to harness the potential of large language models in a production setting while also getting hands-on experience with the emerging field of LLMOps or Big Model Ops.
We deal with a decent volume of support requests in Slack, from the ultra-specific involving on-prem deployments to beginners just getting started. We estimate that allowing our community to freely query the documentation with their questions will save us 10% of our time. We also think it will be extremely helpful for those in the early days of their journey who won’t necessarily be familiar with every page of our documentation and who just want to get moving forward and to try things out. Perhaps most importantly, we hope that the bot will help new ZenML users and community members to feel empowered to ask all the questions they need, without any feeling like they’re taking time from someone else.
What we learned
Evolving Best Practices: In the rapidly evolving world of LLMOps, there are no hard and fast rules yet. Best practices are still being defined as the field advances and professionals gain more experience with large language models. This necessitates flexibility and adaptability when working on such projects.
Rapid Development of Frameworks and Tooling: The pace of development in the field is extremely fast, with frameworks and tools being built and improved continuously. Community contributions to libraries like langchain and llama_index have been instrumental in driving progress. However, it’s essential to recognize that the LLM world still has a long way to go before reaching ‘mature’ tooling.
Data Remains King: As with most machine learning applications, data plays a crucial role in the success of LLMs. For instance, importing Slack support conversations into the vector store led to less robust and useful responses from the Slack bot. Ensuring the quality and relevance of the data used for training and fine-tuning LLMs is vital.
Efficient Querying: There is significant scope for developing smart and efficient ways to query data in LLM applications. It is important to minimize the amount of data sent to external endpoints, such as OpenAI’s, to reduce costs associated with token usage and to maintain data privacy by avoiding exposing real user data on third-party infrastructure.
Adapting Deployment Best Practices: While there are existing best practices for deploying text-based tooling, the growth of LLM usage will likely prompt the need to learn and adapt these practices to accommodate new experiments and innovations. As more professionals explore the potential of LLMs, the collective knowledge and experience will contribute to refining best practices and guiding future developments in the field.
We benefited from many of ZenML’s core features during this project:
- caching allowed us to run our pipelines without needing to wait excessively to repeatedly query the same (mostly-unchanging) documents or data sources
- automatic data versioning for free! (ZenML handles all of this for you out of the box))
- switching between local stacks and deployed cloud infrastructure was simple, as always!
- using ZenML’s post-execution workflow we were able to query the artifact store and get the de-serialised FAISS vector store. This we then used within our deployed slack bot app.
- collaboration through the dashboard: colleagues were able to inspect the progress of our pipelines as well as see where the underlying data was stored through the ZenML dashboard.
Where to go from here?
If you’d like to see our Slack bot in action, feel free to join us in Slack and get help on your problems even when the core team are asleep! Just mention @zenml-bot and you’ll be on your way! The code behind the project is all available in our open-source Projects repository, as usual, and we’d invite you to check out how this all works under the hood.
