Vector similarity search

Qdrant stores embeddings as points and returns the nearest ones to a query vector, and a Beam agent uses that for lookups your workflow needs. Given an incoming item, the agent computes or receives its embedding, queries Qdrant for the closest matches under your score threshold, and reads the payloads attached to each hit. It writes the chosen result back to the record that asked, such as a suggested answer or a similar case. Queries that return nothing above the threshold, or ties too close to call, are handed to a person instead of a weak guess.

Embedding upserts

Qdrant keeps its index current only if new vectors are added as your data grows, and a Beam agent handles that intake. When a fresh document or record appears, the agent reads its content, generates or collects the embedding, and writes the point into the right Qdrant collection with the metadata your schema defines. Existing items are updated in place so duplicates do not accumulate. A record the agent cannot embed cleanly, or one missing a field your payload requires, is set aside for a person rather than indexed in a broken state.

Payload filtering rules

Qdrant pairs each vector with a metadata payload, letting searches filter on real attributes, and a Beam agent applies those filters per request. Reading the context of a query, the agent builds the filter your rule specifies, such as a tenant, date range, or status, and asks Qdrant for matches that satisfy both similarity and the filter. Clean results flow back to the caller. A request whose filter would return an empty or suspiciously large set, or that references a field not in the payload, is paused and referred to a person to check the query.