Each act, each occasion, is worse than the last, but only a little worse. You wait for the next and the next. You wait for one great shocking occasion, thinking that others, when such a shock comes, will join with you in resisting somehow.

If you're doing RAG for finance you need to understand the limits of vector search. Specifically: 📊 They don't generalize out of the training data distribution. 👨‍🍳 The training data mix prioritizes semantics over entities. In other words: If your data contains concepts or named entities that were not in the data used to train that embedding model your retrieval will suffer. And, what's worse is that your retrieval will sooner return info about a similar event at the wrong company than a dissimilar event at the correct company. This is a major source of hallucinations! The best solution we have found is Faceted Srarch. Faceted Search is widely used in Ecommerce search engines. It involves these steps: 🗂️ Tagging documents in your index with relevant labels. 🔖 Tagging the user's search query with those same labels. ✳️ Filtering down to documents that match your query. 🔎 Sarching for the most relevant documents within that set. The facets we have are geography, industry, company, etc. But one problem lingered: How do we adapt to change? Every day new orgs come in to existence via formation, mergers, demergers, joint ventures, spin-offs, carve-outs, etc. Any search engine used in finance should be able to identify and understand these events. It should be able to update its own facets. It should be self organizing. We have been working on this issue and we have a solution. If you'd like to learn more, check out the blog.