Answer Engines

Rag.NET ships with five answer engines. All implement IAnswerEngine and produce a string answer from the query and the retrieved source chunks.

ChatAnswerEngine (default)

Included in Rag.NET core, registered automatically. Builds a single prompt from all source chunks and sends one LLM call.

Best for: Queries with a small number of source chunks and typical question-answering.

No registration needed — it is the default when AddRagNet() is called.

MapReduceAnswerEngine

Install: dotnet add package Rag.NET.AnswerEngines

Runs one LLM call per source chunk in parallel (map). Filters "not found" responses. Combines surviving partial answers in a single reduce call.

Best for: Large document sets where each chunk may individually contain part of the answer. More LLM calls than Chat, but scales with number of chunks.

Registration:

services.AddRagNet(rag => rag.UseMapReduceAnswerEngine());

RefineAnswerEngine

Install: dotnet add package Rag.NET.AnswerEngines

Generates an initial answer from the first source chunk, then iteratively refines it with each subsequent chunk. Sequential — not parallelised.

Best for: When answer coherence matters more than throughput, or when chunks must be incorporated in order.

Registration:

services.AddRagNet(rag => rag.UseRefineAnswerEngine());

FlareAnswerEngine

Install: dotnet add package Rag.NET.AnswerEngines

FLARE (Forward-Looking Active Retrieval) generates the answer one sentence at a time. Each sentence is scored by an IConfidenceScorer against the current context; when confidence drops below FlareOptions.ConfidenceThreshold (default 0.6), the engine runs a lookahead retrieval (original query + low-confidence sentence, LookaheadTopK results), merges the fresh sources into the context (deduplicated by document/chunk, max score kept), and regenerates the sentence once. Generation stops when the model signals completion, after MaxSentences (default 15), and mid-generation retrievals are hard-capped at MaxRetrievals (default 3).

The default scorer is SelfAssessmentConfidenceScorer — one small LLM call per sentence that works with every IChatClient and fails open (score 1.0) on any failure. A logprob-based scorer is a documented extension point: implement IConfidenceScorer and assign it to FlareOptions.Scorer.

Lookahead retrievals default to a plain retrieval — HyDE and multi-query expansion are disabled (the lookahead query already embeds a synthetic document, so expanding it again multiplies hidden LLM calls), reranking stays active. To customize, set FlareOptions.LookaheadRetrievalOptions; it is used verbatim, including TopK (in that case LookaheadTopK is ignored).

Best for: Long-form answers and multi-step reasoning where a single up-front retrieval misses information needed mid-answer. Costs roughly 2 LLM calls per sentence (generation + scoring) plus up to MaxRetrievals plain retrievals and one regeneration call each.

Limitations: FLARE does not consult IConversationMemory (unlike Chat/MapReduce/Refine) — routing a call to Flare via the dispatching engine drops the processed conversation history from the prompt. FLARE also always appends its own fragment protocol (roughly 60 words instructing the model to reply with exactly one sentence, or <DONE>) after any RagOptions.SystemPrompt you supply — it cannot be displaced, because a caller instruction written for a complete reply (e.g. "end with exactly this sentence") is actively harmful applied per fragment: the model would satisfy it every call and never emit <DONE>.

Registration:

services.AddRagNet(rag => rag.UseFlare(o =>
{
    o.ConfidenceThreshold = 0.6;
    o.MaxRetrievals = 3;
    o.MaxSentences = 15;
    o.LookaheadTopK = 3;
    // o.Scorer = new MyLogprobScorer(); // custom IConfidenceScorer
}));

RagResponse.Sources reflects the full merged context including mid-generation retrievals; streaming delegates to the non-streaming path for the same reason (no token-incremental streaming).

DispatchingAnswerEngine

Install: dotnet add package Rag.NET.AnswerEngines

Routes to MapReduce, Refine, Flare, or Chat at call time based on RagOptions.SynthesisStrategy. Allows runtime switching without re-registration. Routing to Flare requires UseFlare() to be registered as well (it provides the confidence scorer and options); requesting SynthesisStrategy.Flare without it throws.

Registration:

services.AddRagNet(rag => rag.UseDispatchingAnswerEngine());

Runtime selection:

var result = await pipeline.AskAsync(query, new RagOptions
{
    SynthesisStrategy = SynthesisStrategy.MapReduce
});

SynthesisStrategy values: Default (Chat), MapReduce, Refine, Flare.

Comparison

Engine	LLM calls	Parallelism	Best for
Chat	1	—	Default, small source sets
MapReduce	N + 1	Yes (map phase)	Large doc sets
Refine	N	No	Order-sensitive synthesis
Flare	~2 per sentence + 1 per lookahead (plain retrieval: no HyDE/multi-query)	No	Long-form, multi-step reasoning
Dispatching	Varies	Depends on strategy	Mixed workloads