Retrieval Chat · Trust Addendum

This document supplements the flowscope Security and Trust Overview with the controls specific to flowscope retrieval chat: a chat interface that answers customer-user questions by retrieving from a corpus of the customer's own documents and generating a grounded response.

Read-only by default

Retrieval chat does not write back to any customer-side system. Documents flow into the corpus through the ingestion paths described in the platform overview (inbox, SFTP, object storage, API, direct upload), and responses flow back to the user through the chat interface. There is no separate channel from flowscope into a customer's system of record. If a customer wants flowscope to take actions based on a chat session (for example "draft this in our CRM"), that requires a separately configured workflow product with its own audit and authorization controls described in the workflow addendum.

Retrieval pipeline

When a document is ingested, it is parsed (using Azure Document Intelligence where OCR is required), split into bounded chunks, and embedded through OpenAI. The embeddings are stored alongside the source chunks in the production Azure Database for PostgreSQL cluster, and the original document blob is retained per the platform retention policy.

When a customer-user asks a question, the question is embedded and used to retrieve the most relevant chunks from that customer's corpus only. Retrieval is filtered by tenant at the query layer; one customer's corpus is never returned in answer to another customer's user.

What the model sees

The model receives only the user's question, the retrieved chunks, and a system prompt directing it to answer solely from those chunks. It does not see the full corpus, the embeddings index, or any other customer's data. Each retrieved chunk carries its source document for citation, and the system is designed to decline when no sufficiently relevant content is found, rather than answer from the model's own training data.

Per-document access controls

Where a customer's corpus contains documents with different visibility rules (for example, finance documents only visible to finance staff), those rules are encoded as filters on the retrieval step and are enforced before the model receives any content. A user who lacks visibility on a document will not see chunks from it in the model's context, and their query will not be embedded against it. Per-document access policies are configured at deployment time in collaboration with the customer's admin.

Query log

Every chat query produces a record that captures the user's identity, the question asked, the chunks retrieved, the response returned, and the timestamp. The query log follows the same retention policy as customer data: kept indefinitely for the duration of the contract so customer admins can review who asked what, and held for a 30-day tail after termination before deletion.