Define the unit before the embedding
A retrieval system returns units of information. That unit might be a product, a paragraph, a support resolution, or an event. If the unit is too broad, relevant evidence can be buried inside it. If it is too narrow, it can lose the context needed to interpret it.
Start by writing down what the application should receive after a successful query. Include the identifier, source link, and any fields required to understand the result without a second guess about its origin.
Keep the source relationship intact
A passage should retain the document it came from. A catalog description should retain the product identifier and useful attributes. A decision should retain its date and supporting sources.
These fields are not decorative metadata. They support updates, grouping, citations, and application-level access decisions. Preserve them while preparing the records that your AI Context collection will search.
Choose a collection-based retrieval path
Polygres AI Context Search is the current path for new semantic retrieval workflows. Build around its documented collection model rather than starting a new integration from an older registration example.
Use semantic retrieval for related meaning. Consider grouped results when several passages belong to the same document or entity and the application needs broader coverage rather than repeated fragments from one source.
Inspect results before tuning everything
Test a small set of clear queries and look at the returned records. When a result is poor, distinguish missing source content from a representation problem or a query-strategy problem.
Changing the collection, query, and source chunking at once makes improvement difficult to attribute. Keep a baseline and make one meaningful change at a time, then compare the evidence returned for the same questions.
Explore the next step
Continue with the platform guide. For the current setup and API contract, use the Polygres documentation.