Flag suspicious text before embeddings are created
Pattern checks can warn, quarantine, or reject common injection, impersonation, and evasion text. They add protection. They do not guarantee detection.
Where it runs
Source documents stay under your control. RAGForge does not require them to be stored on a vendor platform. It does create indexes. Those indexes, and the metadata derived from the documents, still need an export and rebuild plan.
Checks, and what they are not
Pattern checks can warn, quarantine, or reject common injection, impersonation, and evasion text. They add protection. They do not guarantee detection.
The client is told to discard the answer collected so far. Rule-based checks run again as tokens arrive.
That is possible only when chat, embeddings, evaluation, and reranking all use approved local endpoints, and that exact setup has been tested.
Custody
RAGForge is licensed and operated in your VPC or data centre. While it is serving queries it does not call a RAGForge service, and it does not check a licence server.
A meeting of a residency rule is not the same as a ban on external processing. If a model, embedding, or reranking call leaves the environment, that call is part of the data path and has to be approved as such.
Access
A user in one workspace cannot retrieve content from another. The restriction is applied when the query is constructed, not by filtering results afterwards.
People sign in with tokens. Systems use hashed API keys with scopes. There are three roles: admin, member, and viewer.
What is copied, and what is not queried live
Sources include file upload, local folders, S3, GitHub, GitLab, SharePoint, Confluence, Google Drive, and Notion. A sync copies content and metadata into your indexes. RAGForge does not query those systems live at answer time, and it does not enforce their permissions as a live check.
If a source overwrites a document and keeps no archive, the index cannot answer what the document said on an earlier date. The safe result is a refusal.
If RAGForge stops operating
Application data uses Postgres, Elasticsearch, and a configured vector store. Leaving still requires an agreed export of metadata and a procedure to rebuild indexes. Embeddings depend on the model and the schema. Changing the embedding model generally means re-embedding the documents.
Source escrow is an option only after a signed agreement names the escrow agent, confirms what is deposited, and defines the release triggers. It is not in place because this page mentions it.
Approval of a deployment still depends on the setup you choose and the security tests you run. What this page does not claim.