Limits
What RAGForge does not do
These are current limits, not a roadmap. A test on your documents is the evidence that matters. A published score on someone else's questions is not.
-
01
It does not reconstruct history the source did not keep
If a document was overwritten and no version was archived, the safe result is a refusal. A model cannot recover the text that applied on an earlier date.
-
02
It does not query source systems live
Connectors copy content and metadata into the indexes you run. A change in SharePoint, Drive, or a repository is invisible until the next sync. Source-system permissions are not applied as a live check at answer time.
-
03
It does not cite every sentence
An answer that is not a refusal carries source citations. The measured claim is that a citation is present. That is not a guarantee that every statement in the answer is supported.
-
04
It does not name the cause of a bad answer by itself
It records the request, the retrieval result, the evidence decision, and a refusal reason when it does not answer. It does not automatically label a failure as routing, retrieval, or generation.
-
05
A quality score does not pull an answer back
Checks after an answer record scores and review flags. They do not block or withdraw an answer that has already been sent. The refusal happens earlier, before the final answer is written, or not at all.
-
06
Pattern checks do not guarantee detection
Ingestion can warn, quarantine, or reject flagged text. A stream can be stopped if a protected marker appears. Neither check is a proof that hostile text will be found.
-
07
Offline use is a tested configuration, not a default
A deployment with no external processing is possible only when chat, embeddings, evaluation, and reranking all use approved local endpoints, and that combination has been tested. Do not treat an untested setup as isolated.
-
08
Changing the embedding model means rebuilding the index
Embeddings depend on the model and the schema. A new embedding model is not a configuration switch. The documents have to be embedded again.
-
09
The evaluation suite does not ship
Running a scored test against your documents is a service. It is not part of the licensed application. You keep the results of a test you commission. You do not receive the suite.
-
10
This site does not quote a price or a benchmark as proof
Licence, support, and capacity are stated in a proposal, not here. Early scores on public question sets are a baseline for us. They are not a competitor comparison, and they are not evidence that the system is ready for your documents.
What to do with this
If one of these limits is the requirement, test it explicitly
A limit that matters to the decision should be a pass rule on the document test, not a sentence discovered after a rollout.