Class PdfChat

Namespace: LMKit.Retrieval

Assembly: LM-Kit.NET.dll

Provides conversational question-answering over PDF documents by combining intelligent document understanding with grounded response generation.

public sealed class PdfChat : IMultiTurnConversation, IConversation, IKVCache, IDisposable

Inheritance: object

PdfChat

Implements: IMultiTurnConversation

IConversation

IKVCache

IDisposable

Inherited Members: object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Examples

using var chat = new PdfChat(chatModel, embeddingModel);

// Enable vision-based document understanding for complex layouts
chat.DocumentVisionParser = new VlmOcr(visionModel);

// Monitor all operations
chat.DocumentImportProgress += (s, e) =>
{
    if (e.Phase == DocumentImportPhase.PageProcessingStarted)
        Console.WriteLine($"Processing page {e.PageIndex + 1}/{e.TotalPages}");
};

chat.CacheAccessed += (s, e) =>
    Console.WriteLine($"Cache {(e.IsHit ? "hit" : "miss")}: {e.DocumentName}");

chat.PassageRetrievalCompleted += (s, e) =>
    Console.WriteLine($"Retrieved {e.RetrievedCount} passages in {e.Elapsed.TotalMilliseconds:F0}ms");

chat.ResponseGenerationStarted += (s, e) =>
    Console.WriteLine($"Generating response ({(e.UsesFullContext ? "full context" : $"{e.PassageCount} passages")})");

// Load one or more documents
var result = await chat.LoadDocumentAsync("report.pdf");
Console.WriteLine($"Loaded {result.DocumentName}: {result.IndexingMode} ({result.TokenCount} tokens)");

// Ask questions - responses are grounded in document content
var response = await chat.SubmitAsync("What are the key findings?");
Console.WriteLine(response.Response.Completion);

Remarks

Unlike simple text extraction, PdfChat interprets both the physical layout (where content appears on the page) and the logical structure (what that content means) of documents. This enables accurate answers from complex materials including multi-column layouts, tables, forms, and scanned pages.

Document preparation is automatic: smaller documents are provided in full to the model for complete context, while larger documents use passage retrieval to inject only the most relevant excerpts per question.

Load documents using LoadDocument(string, DocumentMetadata, CancellationToken) or LoadDocumentAsync(Stream, string, DocumentMetadata, CancellationToken), then ask questions with Submit(string, CancellationToken). Follow-up questions maintain conversation context for natural multi-turn dialogue.

Constructors

PdfChat(LM, LM, IVectorStore, byte[]): Restores a previous PdfChat session from serialized bytes, with an optional vector store.

PdfChat(LM, LM, IVectorStore, int): Initializes a new instance using separate chat and embedding models with an optional vector store for caching.

PdfChat(LM, LM, IVectorStore, string): Restores a previous PdfChat session from a file on disk, with an optional vector store.

PdfChat(LM, LM, byte[]): Restores a previous PdfChat session from serialized bytes.

PdfChat(LM, LM, int): Initializes a new instance using separate chat and embedding models.

PdfChat(LM, LM, string): Restores a previous PdfChat session from a file on disk.

Properties

ChatHistory: Gets the conversation history containing all exchanged messages.

ContextRemainingSpace: Remaining token budget currently available in the context window.

ContextSize: Gets the total token context size available for this conversation.

ContextWindow: Gets or sets the number of neighboring partitions to include around each matched partition for contextual continuity.

ContextualizationOptions: Gets the options that control how follow-up questions are reformulated when QueryGenerationMode is set to Contextual.

DocumentCount: Gets the total number of documents currently loaded.

DocumentProcessingModality: Gets or sets the modality used for processing document content.

DocumentVisionParser: Gets or sets an optional vision-based analyzer for document understanding.

EmbeddingModel: Gets the embedding model used for computing text embeddings during passage retrieval.

FullDocumentCount: Gets the number of documents loaded in full within the context.

FullDocumentTokenBudget: Gets or sets the maximum token budget for full document inclusion.

HasDocuments: Gets whether any documents have been loaded.

HydeOptions: Gets the options that control how hypothetical answers are generated when QueryGenerationMode is set to HypotheticalAnswer.

ImageDetail: Gets or sets the level of detail used when processing images for vision models. Controls the maximum pixel budget allocated to images, which directly affects token consumption and visual fidelity. Default is High.

IncludePageRenderingsInContext: Gets or sets whether page renderings are included alongside retrieved passages during question answering.

MaxRetrievedPassages: Gets or sets the maximum number of passages retrieved per query.

MaximumCompletionTokens: Gets or sets the maximum number of tokens to generate per response.

MaximumRecallTokens

Maximum number of tokens recalled from Memory per turn.

Defaults to ContextSize / 4. The effective value is automatically capped to at most ContextSize / 2.

Memory

Long-term memory store used to recall relevant context across turns.

Assign an AgentMemory implementation to enable retrieval of relevant text partitions. Retrieved snippets are injected as hidden context up to MaximumRecallTokens.

MinRelevanceScore: Gets or sets the minimum relevance score for retrieved passages.

MmrLambda: Gets or sets the Maximal Marginal Relevance (MMR) lambda parameter that controls the balance between relevance and diversity in retrieval results.

Model: Gets the language model used for generating responses.

MultiQueryOptions: Gets the options that control how query variants are generated when QueryGenerationMode is set to MultiQuery.

OcrEngine: Gets or sets an optional OCR engine for extracting text from image-based pages.

PageProcessingMode: Gets or sets how document pages are processed when loading documents.

PassageRetrievalDocumentCount: Gets the number of documents loaded with passage retrieval.

PreferFullDocumentContext: Gets or sets whether small documents should be provided in full to the model.

QueryGenerationMode: Gets or sets the mode used to generate retrieval queries from user input.

ReasoningLevel: Gets or sets the reasoning level used during response generation.

RemainingDocumentTokenBudget: Gets the remaining token budget available for full document inclusion.

RepetitionPenalty: Gets the repetition penalty configuration used to reduce repetitive outputs.

Reranker: Gets or sets the reranker used to refine passage retrieval results.

RetrievalStrategy: Gets or sets the retrieval strategy that controls how candidate partitions are scored during the initial retrieval phase.

SamplingMode: Gets or sets the token sampling strategy for text generation.

Skills

Registry of Agent Skills available to this conversation.

Skills provide modular capabilities with specialized knowledge and workflows, following the Agent Skills specification.

SystemPrompt: Gets or sets the system prompt that defines the assistant's behavior.

ToolPolicy

Per-turn tool-calling policy used by the conversation runtime.

Controls whether tools are allowed, required, disabled, or whether a specific tool must be used on the current turn.

Tools

Registry of model-callable tools available to this conversation.

Register tools before the first user turn so they are advertised to the model. Tool invocation requires a model that supports tool calls.

UsedDocumentTokens: Gets the number of tokens currently consumed by full-context documents.

Methods

ClearDocuments(): Removes all loaded documents and resets the conversation.

ClearHistory(): Clears the conversation history while keeping loaded documents.

Dispose(): Releases all resources used by this instance.

~PdfChat(): Releases unmanaged resources.

LoadDocument(Stream, string, DocumentMetadata, CancellationToken): Loads a PDF document from a stream.

LoadDocument(string, DocumentMetadata, CancellationToken): Loads a PDF document from the specified file path.

LoadDocumentAsync(Stream, string, DocumentMetadata, CancellationToken): Asynchronously loads a PDF document from a stream.

LoadDocumentAsync(string, DocumentMetadata, CancellationToken): Asynchronously loads a PDF document from the specified file path.

RegenerateResponse(CancellationToken): Regenerates the last response using the same context.

RegenerateResponseAsync(CancellationToken): Asynchronously regenerates the last response using the same context.

SaveSession(): Saves the current session state (conversation, documents, and configuration) to a byte array.

SaveSession(string): Saves the current session state to a file on disk.

Submit(string, CancellationToken): Submits a question and returns a response based on the loaded documents.

SubmitAsync(string, CancellationToken): Asynchronously submits a question and returns a response based on the loaded documents.

Events

AfterTextCompletion: Occurs during response generation as text is produced.

AfterToolInvocation: Fired after a tool invocation finishes (or when it was cancelled/errored).

BeforeToolInvocation: Fired before a tool invocation. Handlers may cancel the call.

CacheAccessed: Occurs when the document cache is accessed during loading.

DocumentImportProgress: Occurs during document import to report progress.

MemoryRecall

Fired when one or more memory partitions are recalled for this turn.

Subscribers may inspect the recalled content and optionally cancel injection by setting Cancel to true.

PassageRetrievalCompleted: Occurs when passage retrieval completes for a query.

ResponseGenerationStarted: Occurs when response generation begins.

ToolApprovalRequired: Fired when a tool invocation requires user approval before execution.

Table of Contents