Hey everyone,
Great to see the level of engagement during the Knowledge Buddy webinar, and a big thank you for all the questions that came in!
Also, a shoutout to Kusum, Pratik, and Ibrahim for sharing both the product perspective and the real implementation experience. That mix really helped bring things to life.
For anyone who couldn’t make it (or wants a second pass), you can watch the replay here:
Getting Real Value from Knowledge Buddy: Demo, Best Practices, and the Road Ahead | Pega
There’s a nice balance between what Knowledge Buddy is today, what it actually takes to make it work well in practice, and how the capabilities are evolving.
A few takeaways that stood out to me:
-
Data really is the foundation - not just connecting sources, but making sure content is clean, structured, and meaningful for the use case
-
It’s an iterative process - refining chunking, prompts, and models based on real feedback is where the gains come from
-
Conversational Buddies and integrations with Agents are a game-changer - Our engineers are adding some really amazing capabilities to Knowledge Buddy, from smarter Hybrid Chunking mechanisms to richer Agent Tools to connect to your Buddies.
We also had some strong questions during the call—around LLM selection, vector storage, hybrid chunking, and moving data across environments. I’ll post those below as individual Q&A responses so we can dig into each.
If you’ve watched the replay or are working with Knowledge Buddy already, jump in:
-
What did you take away from the Webinar?
-
Where are you seeing the biggest challenges (data, validation, or model choice)?
-
Anything you’re experimenting with that others could learn from?
Looking forward to hear your thoughts on this!
Kind regards,
Tim
2 Likes
During the Webinar, we received excellent questions from the audience.
Below I’ll capture the summarized questions and responses.
Architecture
Question: How do we select the LLM option (Amazon Bedrock / Google Vertex API etc.)?
Answer: The recommendation is to choose an LLM aligned with your cloud/VPC setup for performance and data locality (e.g. Nova or Claud models through Bedrock on AWS, Gemini models on Google). Start with smaller, faster models (e.g. Haiku-class) and scale up only if needed to balance quality, latency, and cost.
Question: If only chunks are retrieved, are these already summaries?
Answer: No. Chunks are raw segments of source content. The LLM (or agent) generates the final answer using those chunks as context rather than relying on pre-summarized content.
Question: Is hybrid chunking handled by the selected model?
Answer: No, hybrid extraction and chunking are handled by a dedicated service (including structure-aware processing and OCR). The selected LLM in the Knowledge Buddy configuration is used later to generate responses. More details on this mechanism will be published when this functionality becomes available.
Product Availability:
Question: We are on Cloud 3 but don’t have Knowledge Buddy rulesets. Is it included by default?
Answer: Knowledge Buddy isn’t automatically provisioned in all environments, because availability depends on licensing and AI entitlements.
Question: Do we need to purchase a vector database?
Answer: Vector storage, embeddings, and semantic search are handled within the Knowledge Buddy architecture during ingestion, so no separate procurement is required for the Vector Store. Knowledge Buddy is a licensed product, your organization’s account manager can tell you more about this.
Best Practice:
Question: Are there live customer implementations? Any pitfalls or focus areas?
Answer: Yes, multiple organizations are using Pega Knowledge Buddy in production. Successful implementations focus on data readiness (cleanup, structure, metadata and integration capabilities), defining a “golden truth” set of Q&A pairs for validation, continuous iteration based on user feedback and success metrics, and a strong focus on any governance and approval procedures your organization might have in place to sign off on the go-live of Knowledge Buddy.
Capabilities and roadmap:
Question: Are there plans to include images in responses?
Answer: Image support is evolving: near term this includes OCR-based extraction so image content can be used in creating answers, while richer outputs (including images) are on the roadmap but not yet available today.
Question: What happens if I ask multiple questions in a single paragraph with different contexts?
Answer: Knowledge Buddy supports conversational context, but results are most accurate when questions are clear and focused. The Conversational Knowledge Buddy (or any Agent using the Buddy as a tool) can also split up combined queries into separate questions, so if a user request contains multiple questions, the conversational architecture will be able to handle this.