Hiya Builders! I'm wondering what everyone's take is on the effective use of Project Knowledge with Claude in the context of a Lindy Flow. I want to ground my AI responses in relevant context, e.g. Niche or ICP docs. These are "living" docs that are frequently updated. In Claude, I use custom projects with assistant instructions and the documents are attached as Project Knowledge. The successful response rate here is very high. In Lindy, I am searching the Knowledge Base for the same docs, using the same prompts; however, the response quality varies. Sorta like Lindy isn't grounding properly. I've played with the Search Fuzziness and Max Results, but I'm not seeing the same quality of results as in Claude desktop/web client/api. In a different approach, I'm Getting the Contents of the documents and including the BodyText directly in the Lindy Claude prompt. This works much more reliably. This issue isn't specific to any one flow, rather a more foundational issue that I am hoping to gain some perspective on. Best practices for Lindy, etc.. Thanks!
hey Ronnie! thanks for sharing 🙏🏾 yeah for me it’s always a trade-off decision between using a KB or directly fetching data and injecting to the prompt. you’ll notice when fetching from the KB the snippet is limited to a certain size. that’s great to limit context window but can also cut off important information. So whenever I need to make sure the LLM is fully aware of all the info > directly inject. Whenever I have a large amount of unstructured information > I tend to use the KB. Hope this makes sense!
