When the user's wording is far from the corpus.

HydeRetriever asks the LLM to generate the answer it thinks the user wants (a hypothetical document). It then embeds that synthetic answer and uses the resulting vector for retrieval. The hypothetical answer sits in the same semantic neighbourhood as the real ones in the corpus, so recall jumps.

• Best for: short user queries; technical corpora; "how do I X" questions where the answer paragraph reads very differently from the question
• Cost: one extra LLM completion per query (the hypothetical answer)
• Combine with: vector retrieval, reranker

using LMKit.Retrieval;

var hyde = new HydeRetriever(chatModel, embedder, store)
{
    Options = new HydeOptions
    {
        HypotheticalAnswerMaxTokens = 256,
        TopK = 20,
    }
};

var hits = await hyde.RetrieveAsync("how do I cancel");

MultiQueryRetriever asks the LLM for N rewrites of the original query (rephrasings, synonyms, different angles). It runs retrieval for each rewrite, deduplicates the hits, and returns the merged set. Recall scales sublinearly with N; precision tends to hold or improve.

• Best for: ambiguous queries with multiple valid interpretations
• Cost: one LLM completion to generate N rewrites + N retrieval calls
• Combine with: hybrid retrieval, reranker (essential when N is large)

var mq = new MultiQueryRetriever(chatModel, baseRetriever)
{
    Options = new MultiQueryOptions
    {
        QueryCount     = 4,    // generate 4 rewrites
        TopK           = 10,   // per rewrite
        DeduplicateOnId = true,
    }
};

var hits = await mq.RetrieveAsync("is the trial renewable");

QueryContextualizer rewrites the current turn of a multi-turn conversation into a standalone query that the retriever can use without seeing the earlier history. Essential when the user says "what about for enterprise?" after a paragraph about pricing.

• Best for: conversational RAG (RagChat, PdfChat) with follow-up questions
• Cost: one LLM completion per turn
• Combine with: any retriever; works as a preprocessor

var ctx = new QueryContextualizer(chatModel)
{
    Options = new QueryContextualizationOptions
    {
        IncludeLastNTurns = 4,
    }
};

// User turn N: "what about for enterprise?"
// History     : ["pricing for the team plan", "answer paragraph..."]
string standalone = await ctx.ContextualiseAsync(history, "what about for enterprise?");
// -> "What is the pricing for the enterprise plan?"

// Now feed the standalone query into any retriever.
var hits = await store.SearchAsync(embedder.GetEmbeddings(standalone), topK: 5);