Need Architecture Help: Hitting a wall with RAG Semantic Accuracy vs. Metadata Control Hi everyone 👋 Building a "Deep Search" RAG agent across 3,000+ articles. Goals: <10s latency, strict metadata retention (for UI carousels), and high semantic accuracy. We’ve tested 4 architectures and are currently blocked by the native semantic retrieval quality: - Path 1 (Autonomous Node + Full Table): 30s latency. Text noise kills accuracy. - Path 2 (AI Keyword -> Table Search -> AI Rerank): ~15s latency. Keyword extraction is too rigid and misses semantic intent. - Path 3 (Pointer KB -> Fetch Table Row): Created a KB with only summaries to eliminate noise. Result: KB retrieval accuracy is still surprisingly poor. - Path 4 (Execute Code

findTableRows

+

searchType: "hybrid"

): Searched against

articleSummary

only. Result: Fails to connect obvious concepts (e.g., the query "reflexion capacity" completely misses a summary explicitly discussing "reduces cerebral activity"). Question: It feels like

searchType: "hybrid"

isn't working as expected (or isn<t a thing in botpress), and both KB and Table Vector Search are struggling with semantic depth. Is there a way to improve the underlying embedding quality natively, or a recommended workaround to get high semantic accuracy while maintaining strict structural control over our metadata?

This sounds more like an embedding and chunking limitation than hybrid search itself. Your summaries are likely too compressed, so semantic relationships are getting lost during retrieval. I’d recommend separating semantic retrieval from metadata handling use richer semantic chunks for vector search, then attach the structured metadata after retrieval/reranking. Adding query expansion and a reranking layer usually improves semantic recall significantly while still keeping strict metadata control and low latency. If you need me to help let's connect @brash-plumber-41216

Thanks Teemah! I've created a front matter column with the meta + summary. Though my summary is 2 sentences. I'll test making those summaries richer to see if it holds up to what we are looking for. Just curious. On the latency side, a 10 sec. max return sounds acheivable in botpress ?

Yes, 10s max latency is generally achievable in Botpress, but it depends heavily on how many steps you’re chaining and where the bottleneck sits (KB search, table fetch, LLM call, or rerank). If you keep retrieval lean and avoid multiple sequential AI nodes, you can stay well within that range. In most cases, the latency issue comes more from architecture flow than Botpress itself. @brash-plumber-41216

That does sound acheivable yes!

Hi Ross, GPT 4.1 at 32 sec. on average (36k tokens.) according to stack trace.

can you try nano?

i just did, but i then loose all my output accuracy including carousel.

You should try using this integration for crawling