LlmExtract 0.1.0

LlmExtract

Reliably extract and deserialize JSON from raw LLM text responses.

Every developer calling OpenAI, Anthropic, Ollama, or any LLM writes the same boilerplate to strip markdown fences, remove preamble text, fix trailing commas, and handle edge cases. LlmExtract does all of that in one call.

The Problem

LLM responses are messy:

Sure! Here's the JSON you requested:

```json
{
    "name": "Alice",
    "age": 30, // trailing comma
}

Let me know if you need anything else!

Your `JsonSerializer.Deserialize<T>()` call throws. Every time.

## The Solution

```csharp
using LlmExtract;

// One line — handles fences, preamble, trailing text, and JSON repair
Person person = LlmJson.Extract<Person>(llmResponseText);

Installation

dotnet add package LlmExtract

API

LlmJson.Extract<T>(text)

Extract and deserialize. Throws ExtractionException on failure.

var person = LlmJson.Extract<Person>(responseText);

LlmJson.TryExtract<T>(text, out result, out error)

Safe version — no exceptions.

if (LlmJson.TryExtract<Person>(text, out var person, out var error))
{
    Console.WriteLine(person.Name);
}
else
{
    Console.WriteLine($"Failed: {error.Reason}");        // NoJsonFound or DeserializationFailed
    Console.WriteLine($"Extracted: {error.ExtractedJson}"); // What we tried to parse
}

LlmJson.ExtractRaw(text)

Get the raw JSON string without deserializing.

string? json = LlmJson.ExtractRaw(text); // null if no JSON found

LlmJson.ExtractAll<T>(text)

Extract multiple JSON objects from one response.

IReadOnlyList<Person> people = LlmJson.ExtractAll<Person>(multiObjectResponse);

LlmJsonAccumulator<T> — Streaming

Accumulate streaming chunks and emit objects as they complete.

var accumulator = new LlmJsonAccumulator<Person>();

foreach (var chunk in streamingResponse)
{
    if (accumulator.TryAdd(chunk, out var person))
        Console.WriteLine($"Got: {person.Name}");
}

What It Handles

Zero Dependencies

• Uses System.Text.Json (built into .NET 8+, NuGet reference for netstandard2.0)
• No external packages
• Multi-target: netstandard2.0 + net8.0

Works With Any LLM

OpenAI, Anthropic Claude, Google Gemini, Ollama, Azure OpenAI, Mistral, Llama.cpp — any provider that returns text responses.

License

MIT