I keep hearing that most RAG systems used by AI agents are "badly designed" because they rely solely on vector embeddings. Would a hybrid search approach that includes keyword matching actually solve the memory recall issues we see in complex project management agents?
3 answers
Absolutely. The problem with pure vector search is that it’s terrible at finding specific identifiers like project IDs, dates, or unique technical terms. In a Project Management context, if your agent needs to find "Project-X99," a vector search might return "Project-X98" because they are semantically similar, which is a disaster. By integrating BM25 or traditional SQL-like keyword filtering, you give the agent a "hard" memory to complement its "soft" semantic memory. This hybrid architecture is what differentiates a production-grade agent from a simple prototype that just guesses based on similarity.
Joshua here. Have you considered using metadata filtering alongside the hybrid search to further narrow down the search space for the agent?
Hybrid search is definitely the way forward. It covers the technical gaps that embeddings miss, especially in data-heavy domains.
I agree with Laura. Brian, most "bad" designs are just incomplete. Adding a lexical search layer fixes about 80% of the retrieval errors instantly.
Joshua, metadata is the secret sauce. If you tag your chunks with "date," "priority," or "stakeholder," the agent can apply filters before the retrieval even starts. This drastically reduces the "noise" and ensures the memory is always relevant to the current task phase.