Class RagOptions

Namespace: Qavren.Edge.Rag

Assembly: Qavren.Edge.Rag.dll

Everything the RAG middleware lets a caller change.

public sealed class RagOptions

Inheritance: object

RagOptions

Inherited Members: object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Properties

AttachSourcesToResponse: Publish the source list and the grounded flag on the response. Default true.

CitationsPrompt: The citation instruction rendered below the numbered sources.

ContextFormatter: Replaces both prompts and the whole block layout.

ContextPrompt: The instruction prefix rendered above the numbered sources.

ContextRole: Default ChatRole.User, and never System: retrieved text is untrusted input and must not reach the model as instruction. MEAI's own IChatClient documentation says the application is responsible for prompt injection unless an implementation documents safeguards; sub-project 4 documents that it has none, and that this role choice plus the explicit instruction prefix is the whole of its posture.

ContinueOnRetrievalFailure: Default true: a retrieval failure is logged and the turn continues ungrounded, marked as such, rather than failing. That is what Agent Framework's TextSearchProvider does, and the difference here is that the failure also reaches IEdgeDiagnostics. False rethrows as RagRetrievalFailed.

EmitCitations: Resolve [n] markers into CitationAnnotations. Default true.

KeywordExtractor: Null uses a stop-worded word split capped at eight terms. Returning empty degrades to pure vector search rather than emitting a malformed FTS MATCH.

MaxCharsPerSource: Per-source clamp, with a visible …(truncated) suffix. Default 1200 characters.

MaxContextTokens: Budget for the retrieved-context block alone, not the whole prompt. Default 1024.

QuerySource: Which messages become the query. Default LastUserMessage.

ShortCircuitOnNoContext: Answer "nothing in the sources matched" without calling the model when retrieval returned nothing. Default true.

ShouldRetrieve: Return false to skip retrieval for a turn - a greeting, a follow-up that needs no sources.

TokenCounter: Null uses a chars/4 estimate. Pass the chat model's own counter - reachable as chatClient.GetService<Tokenizer>() - for exactness against the model that will actually read the block.

Top: How many sources to retrieve. Default 5.

Edit this page