- ElasticSearch: Used for storing embeddings, functioning as a vector database and search engine.
- Kong: Serves as an AI gateway for governance and enforcement.
- Datadog: Act as the centralized monitoring tool for the Chatbot and LLMs.
- LaunchDarkly: Used for release and feature management.
Data Preparation
Internally we use 0365 to store all solution briefs and datasheets. Most of the documents are within gathered within the last 2 years so we didn’t spend the time to identify and clean up redundant, obsolete, and trivial (ROT) data. If you need to find a a solution for cleaning up 0365 data sources, AvePoint Opus can be considered.
We generate embeddings by pointing the data source to the correpsonding sharepoint folder. Originally we use Elasticsearch only as a vector database to store the embeddings. Later we found that the top K results from the similarity search is not good enough as input for LLM. We decided to redo all the embeddings using ELSER V2 and using Elasticsearch for relevance search and results rankings. This provides much better results for LLM input.
Governance
We used Kong as the AI gateway or LLM proxy to connect with different LLM models. Kong provides a loosely coupled way to link the chatbot program with backend LLMs. This setup allows us to implement various types of LLM governance within Kong, such as prompt guards, decorators and request/response transformations. If we need to change a prompt template or swap to a different LLM, this can be easily accomplished in Kong. Metrics like token usage and API response times are captured and sent to Datadog for centralized monitoring.
Monitoring
Datadog is selected for LLM monitoring for our internal Chatbot. Datadog gathers metrics from Kong, the Python application, and the inputs/outputs from LLMs. We can easily monitor critical performance metrics such as LLM API response times, input and output tokens, and more. Soft metrics related to AI governance, such as toxicity, hallucination, and prompt ingestion are also monitored in Datadog. Kong and Datadog work nicely together on AI governance. In our case, Kong is mainly used for enforcement, while Datadog provides observability on the overall LLM performance. Datadog also provides comprehensive information on LLM interactions for audit purposes.
Feedback Loop
We incorporated LaunchDarkly to implement a feedback loop. Feedback is collected directly from the Python program and through native LaunchDarkly integration with Datadog. To gather human feedback on model performance, we implemented a thumbs-up/down mechanism on the chatbot interface. These feedbacks are aggregated in Datadog, and both soft and hard metrics from the LLM are being used to toggle feature flags in the chatbot program. For instance, if an LLM experiences long response times or receives increased negative user feedback over a period of time we can automatically trigger feature flags to disable certain LLMs or swap prompt templates. This prevent costly rollbacks to a previous version of the app.
Currently, we manage prompt templates within LaunchDarkly using AI prompt flags, although this can also be handled in Kong. To compare model performance and Chatbot UI designs, we plan to integrate the program with SSO to support A/B testings.
Cost Considerations
We started with a local LLM model “llama3.1 8B” running on a single GPU. We want to save some money during the development phase. Also, some customers have to use an on-prem LLM due to strict company policies. This gives us and idea on whether a local model is good enough for our use cases. Later we added text models from AWS Bedrock for cost and output quality comparison.
Recently Kong announced a new feature on semantic caching, which aims to reduce LLM processing costs by intelligently caching prompts with similar meanings. We cant’ wait to test this feature in our setup to further decrease LLM expenditures on Bedrock.
Conclusion
There are numerous ways to implement a chatbot; you can host everything within a cloud provider such as AWS or build everything in-house. If you need to fine-tune LLM models without managing the infrastructure, AWS can be a good option. For our use case, a local LLM model running on a single RTX 4070 GPU is more than sufficient. Regardless of where your LLM models are deployed, proper governance and guardrails should be implemented. Looking beyond cost and solution functionality, ensuring that LLM deployments adhere to the principles of responsible AI is crucial.
If you’re interested in learning more, please don’t hesitate to contact us: https://vsceptre.com/contact-us/
