Multi-turn conversation
Full conversation history with automatic context carry-over. Follow-up questions resolve pronouns and references seamlessly.
Solutions · Data Processing · Conversational RAG
Turn any knowledge base into a multi-turn conversation.
Single-shot retrieval misses context. Follow-up questions fail because the
retriever forgets what was asked before. LM-Kit's RagChat
combines multi-turn conversation, advanced query reformulation, and grounded
retrieval in a single class that runs entirely on your hardware.
- Follow-up questions lose context from earlier turns
- Ambiguous queries retrieve irrelevant passages
- RAG pipelines need manual orchestration of retrieval, prompt building, and generation
- Cloud RAG APIs expose sensitive corporate knowledge
Smart query reformulation
Four retrieval strategies: Original, Contextual, MultiQuery, and HyDE. Each one optimized for different query complexity levels.
Tool calling and skills
Register tools, built-in functions, and Agent Skills. Extend Q&A with web search, calculations, or any custom operation.
4
Query modes
Query modes
0
Cloud deps
Cloud deps
6+
Events
Events
Conversational Q&A over any knowledge base.
RagChat is LM-Kit.NET's turnkey solution for multi-turn conversational
question-answering over custom knowledge bases. Unlike document-centric RAG,
which manages the full document lifecycle, RagChat operates on a
pre-populated RagEngine that you own and manage. This makes it ideal
for custom corpora, multi-source knowledge bases, and enterprise data that comes
from heterogeneous systems.
A single SubmitAsync() call orchestrates the full pipeline: query
reformulation, semantic retrieval, prompt construction, and grounded response
generation, all with full conversation history preserved across turns.
Complete ownership of your data. RagChat runs 100% on-device. No
API calls, no cloud dependencies, no data leaving your infrastructure. Compliant
with GDPR, HIPAA, and air-gapped environments by design.
RagChat.cs
// Build your knowledge base var ragEngine = new RagEngine(embeddingModel); ragEngine.ImportText(corporateKnowledge); ragEngine.ImportText(productDocs); // Start a multi-turn conversation using var chat = new RagChat(ragEngine, chatModel); chat.QueryGenerationMode = QueryGenerationMode.Contextual; // Ask questions naturally var r1 = await chat.SubmitAsync("What is our refund policy?"); // Follow-up: context is preserved var r2 = await chat.SubmitAsync("Does that apply to digital products?"); // Access retrieved partitions foreach (var p in r2.RetrievedPartitions) Console.WriteLine($"Source: {p.DataSourceIdentifier}");
Four query generation strategies.
Choose the optimal retrieval strategy for your query complexity. Each mode trades off between speed, recall, and precision.
Mode 01
Original
The user's question is sent directly to the retrieval engine. No reformulation, no overhead. The fastest path from question to answer.
- Zero additional latency
- Best for self-contained questions
- Ideal for single-turn interactions
Best for: Direct, self-contained queries
Mode 02
Contextual
Rewrites the user's follow-up question into a self-contained query using conversation history. Resolves pronouns, ellipsis, and implicit references automatically.
- Pronoun and coreference resolution
- Configurable history depth
- Essential for multi-turn dialogue
Best for: Multi-turn conversations
Mode 03
MultiQuery
Generates multiple query variants and searches independently, then fuses results using Reciprocal Rank Fusion (RRF). Maximizes recall for complex or ambiguous questions.
- Configurable variant count
- RRF result fusion
- Higher recall on ambiguous queries
Best for: Complex, multi-faceted queries
Mode 04
HyDE (Hypothetical Answer)
Generates a hypothetical answer first, then retrieves passages similar to that answer. Bridges the vocabulary gap between questions and documents.
- Bridges query/document vocabulary gap
- Best precision for factual retrieval
- Effective on technical corpora
Best for: Technical, domain-specific questions
What happens when you call SubmitAsync.
RagChat orchestrates a five-stage pipeline for every question. Each
stage is observable via events and fully configurable.
Stage 01
Contextualize
Follow-up questions are rewritten into self-contained queries using conversation history.
Stage 02
Retrieve
Semantic search across your RagEngine's data sources. Respects MinRelevanceScore and MaxRetrievedPartitions.
Stage 03
Rank & order
Results ordered by source, section, and partition index. Optional reranking refines relevance scores.
Stage 04
Build prompt
Retrieved context injected into your PromptTemplate via @context and @query placeholders.
Stage 05
Generate
Grounded response generation with streaming via AfterTextCompletion. Supports tool calls, skills, and memory.
Beyond simple Q&A.
RagChat implements IMultiTurnConversation, giving it
the full power of an AI agent combined with grounded retrieval.
Tool calling & skills
Extend RAG with actions
Register custom tools, built-in tools, and Agent Skills that the model can invoke during conversation. Combine knowledge retrieval with live computations, web searches, database lookups, or any external operation.
Tools
ToolRegistry integration
Register any ITool implementation. The model decides when and how to invoke tools during RAG conversations.
Skills
Agent Skills
Define complex capabilities via SKILL.md files. Skills combine system prompts, tools, and behavioral rules into reusable packages.
Approval
Tool approval workflow
ToolApprovalRequired event for human-in-the-loop control. Approve or deny tool invocations before they execute.
Persistent context and full visibility.
Connect AgentMemory for long-term knowledge that persists across
sessions. Every stage of the pipeline emits events for complete observability.
Memory
AgentMemory
RAG-backed persistent memory that survives across sessions. Recall relevant facts from past conversations automatically.
Events
RetrievalCompleted event
Fires after partition retrieval with full details: query used, partitions found, count requested, and elapsed time.
Streaming
Streaming generation
AfterTextCompletion event streams tokens in real time for responsive user experiences. Build interactive chat UIs effortlessly.
RagChat vs PdfChat.
Both provide conversational RAG, but they serve different use cases. Choose the one that matches your data ownership model.
Custom corpora
RagChat
Operates on a pre-populated RagEngine you manage. Full control over data sources, chunking, and the retrieval pipeline. The caller owns the engine lifecycle.
- Bring your own
RagEngine - Multi-source knowledge bases
- 4 query generation modes
- Tool calling, skills, memory
- Custom chunking and storage
Best for: Custom corpora, enterprise knowledge, multi-source data
Document Q&A
PdfChat
Manages the full document lifecycle: import, chunk, embed, cache, and query. Optimized for quick document Q&A with automatic context management.
- Automatic document lifecycle
- Built-in caching
- Smart context loading
- OCR and VLM integration
- Page-level source references
Best for: Document Q&A, PDF chat, single-document focus
Everything you need for enterprise RAG.
Fine-grained control over every aspect of the retrieval and generation pipeline.
Reranking
Semantic reranking
Cross-encoder rerankers refine initial retrieval for significantly higher precision. Access raw and reranked scores on every partition.
Filtering
Relevance filtering
Set MinRelevanceScore (0.0 to 1.0) and MaxRetrievedPartitions to control quality and quantity of retrieved context.
Templates
Custom prompt templates
Configure how context reaches the model with @context and @query placeholders. Optimize prompts for your domain.
Reasoning
Reasoning control
Adjust ReasoningLevel (None, Medium, High) for models that support extended thinking. Balance quality with latency for your use case.
MMR
MMR diversity
Configure Maximal Marginal Relevance on the underlying RagEngine to balance relevance with diversity in retrieved results.
Privacy
100% on-device
Zero cloud dependencies. Your knowledge base, embeddings, and conversations stay on your hardware. GDPR and HIPAA compliant by design.
Storage
Flexible vector storage
Works with any IVectorStore backend: in-memory, built-in SQLite, Qdrant, or your own custom implementation.
APIs
Async and sync APIs
Both SubmitAsync and Submit methods available. Build responsive UIs with async or use synchronous for batch processing and console applications.
Sessions
Conversation management
Full ChatHistory access. ClearHistory() resets conversation state without affecting the knowledge base. Start fresh conversations at any time.
Core classes.
Complete API documentation for building conversational RAG applications.
RagChat
Multi-turn conversational Q&A over a user-managed RagEngine with tool calling and memory.
RagEngine
Core retrieval engine managing data sources, embeddings, and similarity search.
RagQueryResult
Response container with generated text and retrieved partition references.
QueryGenerationMode
Enum for retrieval strategies: Original, Contextual, MultiQuery, HypotheticalAnswer.
DataSource
Content repository for text partitions and sections with metadata support.
PartitionSimilarity
Retrieved partition with similarity scores, source metadata, and embeddings.
Embedder
Generate text and image embeddings for semantic similarity search.
AgentMemory
RAG-backed persistent memory for long-term context across conversation sessions.
IVectorStore
Interface for pluggable vector storage backends (in-memory, SQLite, Qdrant, custom).
Get started in minutes.
Clone working examples from our GitHub repository and start building conversational RAG applications.
Conversational RAG
Multi-turn RAG conversation using RagChat with query contextualization and streaming responses.
Retrieval quality tuning
Compare query generation modes, relevance thresholds, and reranking strategies to optimize retrieval quality.
Help desk knowledge base
Build a help desk assistant with RagChat, multi-source FAQ data, and real-time response streaming.
PDF chat demo
Conversational Q&A over PDF documents with PdfChat for comparison with RagChat's approach.
Persistent memory assistant
Agent with RAG-backed long-term memory that remembers context across conversation sessions.
All samples on GitHub
Browse the complete collection of RAG and conversational AI samples.
Related concepts.
Explore the techniques that power RagChat's retrieval and generation pipeline.
Build it. Read it. Try it.
Working console demos on GitHub, step-by-step how-to guides on the docs site, and the API reference for the classes used on this page.
Build conversational RAG today.
Turn your knowledge base into an intelligent, multi-turn conversation. One NuGet package, zero cloud dependencies, complete data ownership.