Semantic Search Explained: Vector Embeddings, pgvector and FileMaker

Semantic search vs keyword search

Traditional search matches exact words: search for "invoice overdue" and you only find records containing those literal terms. Semantic search instead matches by meaning. A search for "client hasn't paid" can surface a record described as "overdue invoice" or "outstanding balance", because the system compares the meaning of the query to the meaning of each record, not just the characters.

How vector embeddings work

Semantic search is built on vector embeddings: a machine learning model converts a piece of text (a quote, a clinical note, a job description) into a list of numbers — a vector — that represents its meaning in a high-dimensional space. Text with similar meaning ends up with vectors that are mathematically close together. Searching then becomes a matter of embedding the search query the same way and finding the nearest vectors in the dataset.

Where pgvector and Supabase fit in

Storing and searching these vectors efficiently needs a database extension built for the job. pgvector is a Postgres extension that stores vector embeddings and performs similarity search directly inside the database. Supabase packages Postgres with authentication, storage and edge functions, which makes it a practical place to run pgvector alongside the rest of an application's backend — without needing a separate specialised vector database.

Connecting it to FileMaker data

Most businesses running Claris FileMaker do not need to abandon it to get semantic search. A typical pattern is to keep FileMaker as the system of record for day-to-day operations, and sync the text that needs to be searchable — quotes, job cards, notes, correspondence — to Supabase via the FileMaker Data API. Supabase generates and stores the embeddings, handles the similarity search, and the results are surfaced back into the FileMaker interface or a companion web view, complete with a link back to the original record.

Retrieval-augmented assistants extend this pattern one step further: instead of just returning a list of matching records, the system retrieves the most relevant records and uses them as grounding for an answer, with citations back to the source — so staff can verify anything the assistant reports.

Conclusion

Semantic search does not require replacing an existing FileMaker system — it requires a sync path to a vector-capable database and a thoughtful integration back into the tools your team already uses. Done well, it turns years of accumulated notes, quotes and correspondence into something staff can actually search by meaning.