Retrieval pipelines

A Vext pipeline pulls relevant chunks from your indexed data to answer a query. When a question arrives from a connected app, a Beam agent passes it to the pipeline, reads the retrieved context and generated reply, and applies your rule for how confident an answer must be before it is used. It writes the answer and its sources back to the requesting record. Queries where retrieval returns little or conflicting context are routed to a person, so a thin or uncertain answer is reviewed before it reaches a customer.

Grounded answer generation

This produces replies tied to your own documents rather than open guesses. A Beam agent submits the prompt to Vext, reads the grounded response along with the passages it cited, and applies your rule to check the answer actually references your sources. Approved answers are written back to the ticket or record that asked. Responses that cite nothing, or that the confidence signal marks as weak, are handed to a person, who confirms the reply before it is sent, since a wrong grounded answer still reads as authoritative to the reader.

Knowledge base refresh

A pipeline is only as current as the data behind it. On a schedule or a content change, a Beam agent gathers the updated documents, applies your rule for what belongs in the index, and pushes them into Vext so future answers reflect the latest material. It records which documents were added or replaced. Files that look malformed, or updates that would remove a large share of the existing index at once, are flagged for a person, since a bad refresh can quietly degrade every answer the pipeline gives afterward.