RAG Types

RAG Types

The Strategos.Rag package provides vector store adapters for Retrieval-Augmented Generation (RAG) patterns.

IVectorSearchAdapter

Interface for vector similarity search operations.

Methods

Method	Parameters	Returns	Description
SearchAsync	string query, int topK, double minRelevance, CancellationToken ct	Task<IReadOnlyList<SearchResult>>	Searches for similar documents
SearchAsync	float[] embedding, int topK, double minRelevance, CancellationToken ct	Task<IReadOnlyList<SearchResult>>	Searches with pre-computed embedding
UpsertAsync	Document document, CancellationToken ct	Task	Inserts or updates document
DeleteAsync	string documentId, CancellationToken ct	Task	Removes document from index

Example

public class RetrieveContextStep : IWorkflowStep<QueryState>
{
    private readonly IVectorSearchAdapter _vectorSearch;

    public RetrieveContextStep(IVectorSearchAdapter vectorSearch)
    {
        _vectorSearch = vectorSearch;
    }

    public async Task<StepResult<QueryState>> ExecuteAsync(
        QueryState state,
        StepContext context,
        CancellationToken ct)
    {
        var results = await _vectorSearch.SearchAsync(
            query: state.Query,
            topK: 10,
            minRelevance: 0.7,
            ct);

        return state
            .With(s => s.RetrievedContext, results)
            .AsResult();
    }
}

---

SearchResult

Result from a vector similarity search.

Properties

Property	Type	Description
DocumentId	string	Unique identifier of matched document
Content	string	Document text content
Score	double	Similarity score (0.0 to 1.0)
Metadata	Dictionary<string, object>	Additional document metadata

Example

var results = await _vectorSearch.SearchAsync(query, topK: 5, minRelevance: 0.7, ct);

foreach (var result in results)
{
    Console.WriteLine($"ID: {result.DocumentId}");
    Console.WriteLine($"Score: {result.Score:F3}");
    Console.WriteLine($"Content: {result.Content.Substring(0, 100)}...");

    if (result.Metadata.TryGetValue("source", out var source))
    {
        Console.WriteLine($"Source: {source}");
    }
}

---

Document

Represents a document for indexing.

Properties

Property	Type	Description
Id	string	Unique document identifier
Content	string	Text content to embed
Embedding	float[]?	Pre-computed embedding (optional)
Metadata	Dictionary<string, object>	Additional metadata

Example

var document = new Document
{
    Id = "doc-123",
    Content = "Agentic workflows combine deterministic state machines with AI agents...",
    Metadata = new Dictionary<string, object>
    {
        ["source"] = "documentation",
        ["chapter"] = "introduction",
        ["created"] = DateTimeOffset.UtcNow
    }
};

await _vectorSearch.UpsertAsync(document, ct);

---

Implemented Adapters

InMemoryVectorSearchAdapter

In-memory adapter for development and testing.

Aspect	Value
Status	Available
Use Case	Development, unit testing, prototyping
Persistence	None (data lost on restart)

Configuration

services.AddScoped<IVectorSearchAdapter, InMemoryVectorSearchAdapter>();

Limitations

• No persistence between restarts
• Not suitable for production
• Limited scalability

---

Planned Adapters

PgVectorAdapter

PostgreSQL with pgvector extension.

Aspect	Value
Status	Planned
Use Case	Production with existing PostgreSQL
Dependencies	PostgreSQL with pgvector extension

Planned Features

• Leverages existing PostgreSQL infrastructure
• HNSW and IVFFlat index support
• Transactional consistency with workflow state

Planned Configuration

services.AddVectorSearch(options => options
    .UsePgVector(connectionString)
    .WithDimension(1536)
    .WithIndexType(IndexType.HNSW));

---

AzureAISearchAdapter

Azure AI Search for enterprise-scale deployments.

Aspect	Value
Status	Planned
Use Case	Enterprise production, Azure ecosystem
Dependencies	Azure AI Search service

Planned Features

• Managed infrastructure
• Hybrid search (vector + keyword)
• Semantic ranking

Planned Configuration

services.AddVectorSearch(options => options
    .UseAzureAISearch(
        endpoint: "https://your-search.search.windows.net",
        apiKey: "your-api-key",
        indexName: "documents"));

---

RAG Workflow Patterns

Basic Retrieval

Workflow<QueryState>.Create("simple-rag")
    .StartWith<RetrieveContextStep>()
    .Then<GenerateResponseStep>()
    .Finally<FormatOutputStep>();

Retrieval with Reranking

Workflow<QueryState>.Create("reranked-rag")
    .StartWith<RetrieveContextStep>()
    .Then<RerankResultsStep>()
    .Then<GenerateResponseStep>()
    .Finally<FormatOutputStep>();

Multi-Query Retrieval

Workflow<QueryState>.Create("multi-query-rag")
    .StartWith<GenerateQueriesStep>()
    .Fork(
        path => path.Then<RetrieveForQuery1Step>(),
        path => path.Then<RetrieveForQuery2Step>(),
        path => path.Then<RetrieveForQuery3Step>())
    .Join<MergeResultsStep>()
    .Then<DeduplicateStep>()
    .Then<GenerateResponseStep>()
    .Finally<FormatOutputStep>();

---

Example: Complete RAG Step

public class RagQueryStep : IAgentStep<QueryState>
{
    private readonly IVectorSearchAdapter _vectorSearch;
    private readonly IChatClient _chatClient;

    public RagQueryStep(
        IVectorSearchAdapter vectorSearch,
        IChatClient chatClient)
    {
        _vectorSearch = vectorSearch;
        _chatClient = chatClient;
    }

    public async Task<StepResult<QueryState>> ExecuteAsync(
        QueryState state,
        AgentStepContext context,
        CancellationToken ct)
    {
        // Retrieve relevant context
        var searchResults = await _vectorSearch.SearchAsync(
            state.Query,
            topK: 5,
            minRelevance: 0.7,
            ct);

        // Build context string
        var contextText = string.Join("\n\n",
            searchResults.Select(r => r.Content));

        // Generate response with context
        var prompt = $"""
            Answer the following question using only the provided context.

            Context:
            {contextText}

            Question: {state.Query}

            Answer:
            """;

        var response = await _chatClient.GetResponseAsync(prompt, ct);

        return state
            .With(s => s.RetrievedContext, searchResults.ToList())
            .With(s => s.Response, response)
            .AsResult();
    }
}

---

Registration

// Development
services.AddScoped<IVectorSearchAdapter, InMemoryVectorSearchAdapter>();

// Production (when available)
services.AddVectorSearch(options => options
    .UsePgVector(connectionString)
    .WithEmbeddingClient<OpenAIEmbeddingClient>());