llm.rb 4.8.0 → 4.10.0

data/README.md

@@ -4,751 +4,544 @@
 <p align="center">
   <a href="https://0x1eef.github.io/x/llm.rb?rebuild=1"><img src="https://img.shields.io/badge/docs-0x1eef.github.io-blue.svg" alt="RubyDoc"></a>
   <a href="https://opensource.org/license/0bsd"><img src="https://img.shields.io/badge/License-0BSD-orange.svg?" alt="License"></a>
-  <a href="https://github.com/llmrb/llm.rb/tags"><img src="https://img.shields.io/badge/version-4.8.0-green.svg?" alt="Version"></a>
+  <a href="https://github.com/llmrb/llm.rb/tags"><img src="https://img.shields.io/badge/version-4.10.0-green.svg?" alt="Version"></a>
 </p>
 
 ## About
 
-llm.rb is a zero-dependency Ruby toolkit for Large Language Models that
-includes OpenAI, Gemini, Anthropic, xAI (Grok), zAI, DeepSeek, Ollama,
-and LlamaCpp. The toolkit includes full support for chat, streaming,
-tool calling, audio, images, files, and structured outputs.
+llm.rb is a Ruby-centric toolkit for building real LLM-powered systems — where
+LLMs are part of your architecture, not just API calls. It gives you explicit
+control over contexts, tools, concurrency, and providers, so you can compose
+reliable, production-ready workflows without hidden abstractions.
 
-And it is licensed under the [0BSD License](https://choosealicense.com/licenses/0bsd/) &ndash;
-one of the most permissive open source licenses, with minimal conditions for use,
-modification, and/or distribution. Attribution is appreciated, but not required
-by the license. Built with [good music](https://www.youtube.com/watch?v=SNvaqwTbn14)
-and a lot of ☕️.
+Built for engineers who want to understand and control their LLM systems. No
+frameworks, no hidden magic — just composable primitives for building real
+applications, from scripts to full systems like [Relay](https://github.com/llmrb/relay).
 
-## Quick start
+Jump to [Quick start](#quick-start), discover its [capabilities](#capabilities), read about
+its [architecture](#architecture--execution-model) or watch the
+[Screencast](https://www.youtube.com/watch?v=x1K4wMeO_QA) for a deep dive into the design
+and capabilities of llm.rb.
 
-#### REPL
+## What Makes It Different
 
-The [LLM::Session](https://0x1eef.github.io/x/llm.rb/LLM/Session.html) class provides
-a session with an LLM provider that maintains conversation history and context across
-multiple requests. The following example implements a simple REPL loop, and the response
-is streamed to the terminal in real-time as it arrives from the provider. The provider
-happens to be OpenAI in this case but it could be any other provider, and `$stdout`
-could be any object that implements the `#<<` method:
+Most LLM libraries stop at requests and responses. <br>
+llm.rb is built around the state and execution model around them:
 
-```ruby
-#!/usr/bin/env ruby
-require "llm"
+- **Contexts are central** <br>
+  They hold history, tools, schema, usage, cost, persistence, and execution state.
+- **Tool execution is explicit** <br>
+  Run local, provider-native, and MCP tools sequentially or concurrently with threads, fibers, or async tasks.
+- **One API across providers and capabilities** <br>
+  The same model covers chat, files, images, audio, embeddings, vector stores, and more.
+- **Thread-safe where it matters** <br>
+  Providers are shareable, while contexts stay isolated and stateful.
+- **Local metadata, fewer extra API calls** <br>
+  A built-in registry provides model capabilities, limits, pricing, and cost estimation.
+- **Stdlib-only by default** <br>
+  llm.rb runs on the Ruby standard library by default, with providers, optional features, and the model registry loaded only when you use them.
 
-llm = LLM.openai(key: ENV["KEY"])
-ses = LLM::Session.new(llm, stream: $stdout)
-loop do
-  print "> "
-  ses.talk(STDIN.gets || break)
-  puts
-end
-```
+## Architecture & Execution Model
+
+llm.rb is built in layers, each providing explicit control:
 
-#### Schema
+```
+┌─────────────────────────────────────────┐
+│          Your Application               │
+├─────────────────────────────────────────┤
+│         Contexts & Agents               │ ← Stateful workflows
+├─────────────────────────────────────────┤
+│           Tools & Functions             │ ← Concurrent execution
+├─────────────────────────────────────────┤
+│   Unified Provider API (OpenAI, etc.)   │ ← Provider abstraction
+├─────────────────────────────────────────┤
+│      HTTP, JSON, Thread Safety          │ ← Infrastructure
+└─────────────────────────────────────────┘
+```
 
-The [LLM::Schema](https://0x1eef.github.io/x/llm.rb/LLM/Schema.html) class provides
-a simple DSL for describing the structure of a response that an LLM emits according
-to a JSON schema. The schema lets a client describe what JSON object an LLM should
-emit, and the LLM will abide by the schema to the best of its ability:
+### Key Design Decisions
+
+- **Thread-safe providers** - `LLM::Provider` instances are safe to share across threads
+- **Thread-local contexts** - `LLM::Context` should generally be kept thread-local
+- **Lazy loading** - Providers, optional features, and the model registry load on demand
+- **JSON adapter system** - Swap JSON libraries (JSON/Oj/Yajl) for performance
+- **Registry system** - Local metadata for model capabilities, limits, and pricing
+- **Provider adaptation** - Normalizes differences between OpenAI, Anthropic, Google, and other providers
+- **Structured tool execution** - Errors are captured and returned as data, not raised unpredictably
+- **Function vs Tool APIs** - Choose between class-based tools and closure-based functions
+
+## Capabilities
+
+llm.rb provides a complete set of primitives for building LLM-powered systems:
+
+- **Chat & Contexts** — stateless and stateful interactions with persistence
+- **Streaming** — real-time responses across providers
+- **Tool Calling** — define and execute functions with automatic orchestration
+- **Concurrent Execution** — threads, async tasks, and fibers
+- **Agents** — reusable, preconfigured assistants with tool auto-execution
+- **Structured Outputs** — JSON schema-based responses
+- **MCP Support** — integrate external tool servers dynamically
+- **Multimodal Inputs** — text, images, audio, documents, URLs
+- **Audio** — text-to-speech, transcription, translation
+- **Images** — generation and editing
+- **Files API** — upload and reference files in prompts
+- **Embeddings** — vector generation for search and RAG
+- **Vector Stores** — OpenAI-based retrieval workflows
+- **Cost Tracking** — estimate usage without API calls
+- **Observability** — tracing, logging, telemetry
+- **Model Registry** — local metadata for capabilities, limits, pricing
+
+## Quick Start
+
+#### Concurrent Tools
+
+llm.rb provides explicit concurrency control for tool execution. The
+`wait(:thread)` method spawns each pending function in its own thread and waits
+for all to complete. You can also use `:fiber` for cooperative multitasking or
+`:task` for async/await patterns (requires the `async` gem). The context
+automatically collects all results and reports them back to the LLM in a
+single turn, maintaining conversation flow while parallelizing independent
+operations:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
-require "pp"
-
-class Report < LLM::Schema
-  property :category, String, "Report category", required: true
-  property :summary, String, "Short summary", required: true
-  property :services, Array[String], "Impacted services", required: true
-  property :timestamp, String, "When it happened", optional: true
-end
 
 llm = LLM.openai(key: ENV["KEY"])
-ses = LLM::Session.new(llm, schema: Report)
-res = ses.talk("Structure this report: 'Database latency spiked at 10:42 UTC, causing 5% request timeouts for 12 minutes.'")
-pp res.content!
+ctx = LLM::Context.new(llm, stream: $stdout, tools: [FetchWeather, FetchNews, FetchStock])
 
-##
-# {
-#   "category" => "Performance Incident",
-#   "summary" => "Database latency spiked, causing 5% request timeouts for 12 minutes.",
-#   "services" => ["Database"],
-#   "timestamp" => "2024-06-05T10:42:00Z"
-# }
+# Execute multiple independent tools concurrently
+ctx.talk("Summarize the weather, headlines, and stock price.")
+ctx.talk(ctx.functions.wait(:thread)) while ctx.functions.any?
 ```
 
-#### Tools
+#### MCP
 
-The [LLM::Tool](https://0x1eef.github.io/x/llm.rb/LLM/Tool.html) class lets you
-define callable tools for the model. Each tool is described to the LLM as a function
-it can invoke to fetch information or perform an action. The model decides when to
-call tools based on the conversation; when it does, llm.rb runs the tool and sends
-the result back on the next request. The following example implements a simple tool
-that runs shell commands:
+llm.rb integrates with the Model Context Protocol (MCP) to dynamically discover
+and use tools from external servers. This example starts a filesystem MCP
+server over stdio and makes its tools available to a context, enabling the LLM
+to interact with the local file system through a standardized interface:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
-class System < LLM::Tool
-  name "system"
-  description "Run a shell command"
-  param :command, String, "Command to execute", required: true
-
-  def call(command:)
-    {success: system(command)}
-  end
-end
-
 llm = LLM.openai(key: ENV["KEY"])
-ses = LLM::Session.new(llm, tools: [System])
-ses.talk("Run `date`.")
-ses.talk(ses.functions.map(&:call)) # report return value to the LLM
+mcp = LLM.mcp(stdio: {argv: ["npx", "-y", "@modelcontextprotocol/server-filesystem", Dir.pwd]})
+
+begin
+  mcp.start
+  ctx = LLM::Context.new(llm, stream: $stdout, tools: mcp.tools)
+  ctx.talk("List the directories in this project.")
+  ctx.talk(ctx.functions.call) while ctx.functions.any?
+ensure
+  mcp.stop
+end
 ```
 
-#### Agents
-
-The [LLM::Agent](https://0x1eef.github.io/x/llm.rb/LLM/Agent.html)
-class provides a class-level DSL for defining reusable, preconfigured
-assistants with defaults for model, tools, schema, and instructions.
-Instructions are injected only on the first request, and unlike
-[LLM::Session](https://0x1eef.github.io/x/llm.rb/LLM/Session.html),
-an [LLM::Agent](https://0x1eef.github.io/x/llm.rb/LLM/Agent.html)
-will automatically call tools when needed:
+You can also connect to a hosted MCP server over HTTP. This is useful when the
+server already runs remotely and exposes MCP through a URL instead of a local
+process:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
-class SystemAdmin < LLM::Agent
-  model "gpt-4.1"
-  instructions "You are a Linux system admin"
-  tools Shell
-  schema Result
-end
-
 llm = LLM.openai(key: ENV["KEY"])
-agent = SystemAdmin.new(llm)
-res = agent.talk("Run 'date'")
+mcp = LLM.mcp(http: {
+  url: "https://api.githubcopilot.com/mcp/",
+  headers: {"Authorization" => "Bearer #{ENV.fetch("GITHUB_PAT")}"}
+})
+
+begin
+  mcp.start
+  ctx = LLM::Context.new(llm, stream: $stdout, tools: mcp.tools)
+  ctx.talk("List the available GitHub MCP toolsets.")
+  ctx.talk(ctx.functions.call) while ctx.functions.any?
+ensure
+  mcp.stop
+end
 ```
 
-#### Prompts
+#### Streaming Chat
 
-The [LLM::Prompt](https://0x1eef.github.io/x/llm.rb/LLM/Prompt.html)
-class represents a single request composed of multiple messages.
-It is useful when a single turn needs more than one message, for example:
-system instructions plus one or more user messages, or a replay of
-prior context:
+This example demonstrates llm.rb's streaming support. The `stream: $stdout`
+parameter tells the context to write responses incrementally as they arrive
+from the LLM. The `Context` object manages the conversation history, and
+`talk()` sends your input while automatically appending both your message and
+the LLM's response to the context. Streams accept any object with `#<<`,
+giving you flexibility to pipe output to files, network sockets, or custom
+buffers:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
 llm = LLM.openai(key: ENV["KEY"])
-ses = LLM::Session.new(llm)
-
-prompt = ses.prompt do
-  system "Be concise and show your reasoning briefly."
-  user "If a train goes 60 mph for 1.5 hours, how far does it travel?"
-  user "Now double the speed for the same time."
+ctx = LLM::Context.new(llm, stream: $stdout)
+loop do
+  print "> "
+  ctx.talk(STDIN.gets || break)
+  puts
 end
-
-ses.talk(prompt)
 ```
 
-But prompts are not session-scoped. [LLM::Prompt](https://0x1eef.github.io/x/llm.rb/LLM/Prompt.html)
-is a first-class object that you can build and pass around independently of a session.
-This enables patterns where you compose a prompt in one part of your code,
-and execute it through a session elsewhere:
+#### Tool Calling
+
+Tools in llm.rb can be defined as classes inheriting from `LLM::Tool` or as
+closures using `LLM.function`. When the LLM requests a tool call, the context
+stores `Function` objects in `ctx.functions`. The `call()` method executes all
+pending functions and returns their results to the LLM. Tools support
+structured parameters with JSON Schema validation and automatically adapt to
+each provider's API format (OpenAI, Anthropic, Google, etc.):
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
-llm = LLM.openai(key: ENV["KEY"])
-ses = LLM::Session.new(llm)
+class System < LLM::Tool
+  name "system"
+  description "Run a shell command"
+  param :command, String, "Command to execute", required: true
 
-prompt = LLM::Prompt.new(llm) do
-  system "Be concise and show your reasoning briefly."
-  user "If a train goes 60 mph for 1.5 hours, how far does it travel?"
-  user "Now double the speed for the same time."
+  def call(command:)
+    {success: system(command)}
+  end
 end
 
-ses.talk(prompt)
+llm = LLM.openai(key: ENV["KEY"])
+ctx = LLM::Context.new(llm, stream: $stdout, tools: [System])
+ctx.talk("Run `date`.")
+ctx.talk(ctx.functions.call) while ctx.functions.any?
 ```
 
-#### Threads
-
-llm.rb is designed for threaded environments with throughput in mind.
-Locks are used selectively, and localized state is preferred wherever
-possible. Blanket locking across every class could help guarantee
-correctness but it could also add contention, reduce throughput,
-and increase complexity.
-
-That's why we decided to optimize for both correctness and throughput
-instead. An important part of that design is guaranteeing that
-[LLM::Provider](https://0x1eef.github.io/x/llm.rb/LLM/Provider.html)
-is safe to share and use across threads. [LLM::Session](https://0x1eef.github.io/x/llm.rb/LLM/Session.html) and
-[LLM::Agent](https://0x1eef.github.io/x/llm.rb/LLM/Agent.html) are
-stateful objects that should be kept local to a single thread.
-
-[LLM::Tracer](https://0x1eef.github.io/x/llm.rb/LLM/Tracer.html) and its
-subclasses are also designed to be thread-local, which means that
-`llm.tracer = ...` only impacts the current thread and must be set
-again in each thread where a tracer is desired. This avoids contention
-on tracer state, keeps tracing isolated per thread, and allows different
-tracers to be used in different threads simultaneously.
-
-So the recommended pattern is to keep one session, tracer or agent per
-thread, and share a provider across multiple threads:
+#### Structured Outputs
 
+The `LLM::Schema` system lets you define JSON schemas that LLMs must follow.
+Schemas can be defined as classes with `property` declarations or built
+programmatically using a fluent interface. When you pass a schema to a context,
+llm.rb automatically configures the provider's JSON mode and validates
+responses against your schema. The `content!` method returns the parsed JSON
+object, while errors are captured as structured data rather than raising
+exceptions:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
+require "pp"
 
-llm = LLM.openai(key: ENV["KEY"]).persist!
-schema = llm.schema.object(answer: llm.schema.integer.required)
+class Report < LLM::Schema
+  property :category, Enum["performance", "security", "outage"], "Report category", required: true
+  property :summary, String, "Short summary", required: true
+  property :impact, OneOf[String, Integer], "Primary impact, as text or a count", required: true
+  property :services, Array[String], "Impacted services", required: true
+  property :timestamp, String, "When it happened", optional: true
+end
 
-vals = 10.times.map do |x|
-  Thread.new do
-    llm.tracer = LLM::Tracer::Logger.new(llm, path: "thread#{x}.log")
-    ses = LLM::Session.new(llm, schema:)
-    res = ses.talk "#{x} + 5 = ?"
-    res.content!
-  end
-end.map(&:value)
+llm = LLM.openai(key: ENV["KEY"])
+ctx = LLM::Context.new(llm, schema: Report)
+res = ctx.talk("Structure this report: 'Database latency spiked at 10:42 UTC, causing 5% request timeouts for 12 minutes.'")
+pp res.content!
 
-vals.each { |val| puts val }
+# {
+#   "category" => "performance",
+#   "summary" => "Database latency spiked, causing 5% request timeouts for 12 minutes.",
+#   "impact" => "5% request timeouts",
+#   "services" => ["Database"],
+#   "timestamp" => "2024-06-05T10:42:00Z"
+# }
 ```
 
-## Features
-
-#### General
-- ✅  Unified API across providers
-- 📦  Zero runtime deps (stdlib-only)
-- 🧵  Thread-safe providers for multi-threaded workloads
-- 🧩  Pluggable JSON adapters (JSON, Oj, Yajl, etc)
-- 🧱  Builtin tracer API ([LLM::Tracer](https://0x1eef.github.io/x/llm.rb/LLM/Tracer.html))
+## Providers
 
-#### Optionals
+llm.rb supports multiple LLM providers with a unified API.
+All providers share the same context, tool, and concurrency interfaces, making
+it easy to switch between cloud and local models:
 
-- ♻️  Optional persistent HTTP pool via net-http-persistent ([net-http-persistent](https://github.com/drbrain/net-http-persistent))
-- 📈  Optional telemetry support via OpenTelemetry ([opentelemetry-sdk](https://github.com/open-telemetry/opentelemetry-ruby))
-- 🪵  Optional logging support via Ruby's standard library ([ruby/logger](https://github.com/ruby/logger))
+- **OpenAI** (`LLM.openai`)
+- **Anthropic** (`LLM.anthropic`)
+- **Google** (`LLM.google`)
+- **DeepSeek** (`LLM.deepseek`)
+- **xAI** (`LLM.xai`)
+- **zAI** (`LLM.zai`)
+- **Ollama** (`LLM.ollama`)
+- **Llama.cpp** (`LLM.llamacpp`)
 
-#### Chat, Agents
-- 🧠  Stateless + stateful chat (completions + responses)
-- 💾  Save and restore sessions across processes
-- 🤖  Tool calling / function execution
-- 🔁  Agent tool-call auto-execution (bounded)
-- 🗂️  JSON Schema structured output
-- 📡  Streaming responses
-
-#### Media
-- 🗣️  TTS, transcription, translation
-- 🖼️  Image generation + editing
-- 📎  Files API + prompt-aware file inputs
-- 📦  Streaming multipart uploads (no full buffering)
-- 💡  Multimodal prompts (text, documents, audio, images, video, URLs)
-
-#### Embeddings
-- 🧮  Embeddings
-- 🧱  OpenAI vector stores (RAG)
+## Production
 
-#### Miscellaneous
-- 📜  Models API
-- 🔧  OpenAI responses + moderations
+#### Ready for production
 
-## Matrix
+llm.rb is designed for production use from the ground up:
 
-| Feature / Provider                  | OpenAI | Anthropic | Gemini | DeepSeek | xAI (Grok) | zAI    | Ollama | LlamaCpp |
-|--------------------------------------|:------:|:---------:|:------:|:--------:|:----------:|:------:|:------:|:--------:|
-| **Chat Completions**                 | ✅     | ✅        | ✅     | ✅       | ✅         | ✅     | ✅     | ✅       |
-| **Streaming**                        | ✅     | ✅        | ✅     | ✅       | ✅         | ✅     | ✅     | ✅       |
-| **Tool Calling**                     | ✅     | ✅        | ✅     | ✅       | ✅         | ✅     | ✅     | ✅       |
-| **JSON Schema / Structured Output**  | ✅     | ❌        | ✅     | ❌       | ✅         | ❌     | ✅*    | ✅*      |
-| **Embeddings**                       | ✅     | ✅        | ✅     | ✅       | ❌         | ❌     | ✅     | ✅       |
-| **Multimodal Prompts** *(text, documents, audio, images, videos, URLs, etc)* | ✅     | ✅        | ✅     | ✅       | ✅         | ❌     | ✅     | ✅       |
-| **Files API**                        | ✅     | ✅        | ✅     | ❌       | ❌         | ❌     | ❌     | ❌       |
-| **Models API**                       | ✅     | ✅        | ✅     | ✅       | ✅         | ❌     | ✅     | ✅       |
-| **Audio (TTS / Transcribe / Translate)** | ✅  | ❌        | ✅     | ❌       | ❌         | ❌     | ❌     | ❌       |
-| **Image Generation & Editing**       | ✅     | ❌        | ✅     | ❌       | ✅         | ❌     | ❌     | ❌       |
-| **Local Model Support**              | ❌     | ❌        | ❌     | ❌       | ❌         | ❌     | ✅     | ✅       |
-| **Vector Stores (RAG)**               | ✅     | ❌        | ❌     | ❌       | ❌         | ❌     | ❌     | ❌       |
-| **Responses**                        | ✅     | ❌        | ❌     | ❌       | ❌         | ❌     | ❌     | ❌       |
-| **Moderations**                      | ✅     | ❌        | ❌     | ❌       | ❌         | ❌     | ❌     | ❌       |
+- **Thread-safe providers** - Share `LLM::Provider` instances across your application
+- **Thread-local contexts** - Keep `LLM::Context` instances thread-local for state isolation
+- **Cost tracking** - Know your spend before the bill arrives
+- **Observability** - Built-in tracing with OpenTelemetry support
+- **Persistence** - Save and restore contexts across processes
+- **Performance** - Swap JSON adapters and enable HTTP connection pooling
+- **Error handling** - Structured errors, not unpredictable exceptions
 
-\* JSON Schema support in Ollama/LlamaCpp depends on the model, not the API.
+#### Tracing
 
+llm.rb includes built-in tracers for local logging, OpenTelemetry, and
+LangSmith. Assign a tracer to a provider and all context requests and tool
+calls made through that provider will be instrumented. Tracers are local to
+the current fiber, so the same provider can use different tracers in different
+concurrent tasks without interfering with each other.
 
-## Examples
-
-### Providers
-
-#### LLM::Provider
-
-All providers inherit from [LLM::Provider](https://0x1eef.github.io/x/llm.rb/LLM/Provider.html) &ndash;
-they share a common interface and set of functionality. Each provider can be instantiated
-using an API key (if required) and an optional set of configuration options via
-[the singleton methods of LLM](https://0x1eef.github.io/x/llm.rb/LLM.html). For example:
-
-```ruby
-#!/usr/bin/env ruby
-require "llm"
-
-##
-# remote providers
-llm = LLM.openai(key: "yourapikey")
-llm = LLM.gemini(key: "yourapikey")
-llm = LLM.anthropic(key: "yourapikey")
-llm = LLM.xai(key: "yourapikey")
-llm = LLM.zai(key: "yourapikey")
-llm = LLM.deepseek(key: "yourapikey")
-
-##
-# local providers
-llm = LLM.ollama(key: nil)
-llm = LLM.llamacpp(key: nil)
-```
-
-#### LLM::Response
-
-All provider methods that perform requests return an
-[LLM::Response](https://0x1eef.github.io/x/llm.rb/LLM/Response.html).
-If the HTTP response is JSON (`content-type: application/json`),
-`response.body` is parsed into an
-[LLM::Object](https://0x1eef.github.io/x/llm.rb/LLM/Object.html) for
-dot-access. For non-JSON responses, `response.body` is a raw string.
-It is also possible to access top-level keys directly on the response
-(eg: `res.object` instead of `res.body.object`):
+Use the logger tracer when you want structured logs through Ruby's standard
+library:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
 llm = LLM.openai(key: ENV["KEY"])
-res = llm.models.all
-puts res.object
-puts res.data.first.id
-```
-
-#### Persistence
-
-The llm.rb library can maintain a process-wide connection pool
-for each provider that is instantiated. This feature can improve
-performance but it is optional, the implementation depends on
-[net-http-persistent](https://github.com/drbrain/net-http-persistent),
-and the gem should be installed separately:
-
-```ruby
-#!/usr/bin/env ruby
-require "llm"
+llm.tracer = LLM::Tracer::Logger.new(llm, io: $stdout)
 
-llm  = LLM.openai(key: ENV["KEY"]).persist!
-res1 = llm.responses.create "message 1"
-res2 = llm.responses.create "message 2", previous_response_id: res1.response_id
-res3 = llm.responses.create "message 3", previous_response_id: res2.response_id
-puts res3.output_text
+ctx = LLM::Context.new(llm)
+ctx.talk("Hello")
 ```
 
-#### Telemetry
-
-The llm.rb library includes telemetry support through its tracer API, and it
-can be used to trace LLM requests. It can be useful for debugging, monitoring,
-and observability. The primary use case in mind is integration with tools like
-[LangSmith](https://www.langsmith.com/).
-
-It is worth mentioning that tracers are local to a thread, and they
-should be configured per thread. That means that `llm.tracer = LLM::Tracer::Telemetry.new(llm)`
-only impacts the current thread, and it should be repeated in each thread where
-tracing is required.
-
-The telemetry implementation uses the [opentelemetry-sdk](https://github.com/open-telemetry/opentelemetry-ruby)
-and is based on the [gen-ai telemetry spec(s)](https://github.com/open-telemetry/semantic-conventions/blob/main/docs/gen-ai/).
-This feature is optional, disabled by default, and the [opentelemetry-sdk](https://github.com/open-telemetry/opentelemetry-ruby)
-gem should be installed separately. Please also note that llm.rb will take care of
-loading and configuring the [opentelemetry-sdk](https://github.com/open-telemetry/opentelemetry-ruby)
-library for you, and llm.rb configures an in-memory exporter that doesn't have
-external dependencies by default:
+Use the telemetry tracer when you want OpenTelemetry spans. This requires the
+`opentelemetry-sdk` gem, and exporters such as OTLP can be added separately:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
-require "pp"
 
 llm = LLM.openai(key: ENV["KEY"])
 llm.tracer = LLM::Tracer::Telemetry.new(llm)
 
-ses = LLM::Session.new(llm)
-ses.talk "Hello world!"
-ses.talk "Adios."
-ses.tracer.spans.each { |span| pp span }
+ctx = LLM::Context.new(llm)
+ctx.talk("Hello")
+pp llm.tracer.spans
 ```
 
-The llm.rb library also supports export through the OpenTelemetry Protocol (OTLP).
-OTLP is a standard protocol for exporting telemetry data, and it is supported by
-multiple observability tools. By default the export is batched in the background,
-and happens automatically but short lived scripts might need to
-[explicitly flush](https://0x1eef.github.io/x/llm.rb/LLM/Tracer/Telemetry#flush!-instance_method)
-the exporter before they exit &ndash; otherwise some telemetry data could be lost:
-
-```ruby
- #!/usr/bin/env ruby
- require "llm"
- require "opentelemetry-exporter-otlp"
-
- endpoint = "https://api.smith.langchain.com/otel/v1/traces"
- exporter = OpenTelemetry::Exporter::OTLP::Exporter.new(endpoint:)
-
- llm = LLM.openai(key: ENV["KEY"])
- llm.tracer = LLM::Tracer::Telemetry.new(llm, exporter:)
-
- ses = LLM::Session.new(llm)
- ses.talk "hello"
- ses.talk "how are you?"
-
- at_exit do
-   # Helpful for short-lived scripts, otherwise the exporter
-   # might not have time to flush pending telemetry data
-   ses.tracer.flush!
- end
- ```
-
-#### Logger
-
-The llm.rb library includes simple logging support through its
-tracer API, and Ruby's standard library ([ruby/logger](https://github.com/ruby/logger)).
-This feature is optional, disabled by default, and it can be useful for debugging and/or
-monitoring requests to LLM providers. The `path` or `io` options can be used to choose
-where logs are written, and by default it is set to `$stdout`. Like other tracers,
-the logger tracer is local to a thread:
+Use the LangSmith tracer when you want LangSmith-compatible metadata and trace
+grouping on top of the telemetry tracer:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
 llm = LLM.openai(key: ENV["KEY"])
-llm.tracer = LLM::Tracer::Logger.new(llm, io: $stdout)
+llm.tracer = LLM::Tracer::Langsmith.new(
+  llm,
+  metadata: {env: "dev"},
+  tags: ["chatbot"]
+)
 
-ses = LLM::Session.new(llm)
-ses.talk "Hello world!"
-ses.talk "Adios."
+ctx = LLM::Context.new(llm)
+ctx.talk("Hello")
 ```
 
-#### Serialization
+#### Thread Safety
 
-[LLM::Session](https://0x1eef.github.io/x/llm.rb/LLM/Session.html) can be
-serialized and deserialized across process boundaries and persisted to
-storage such as files, a `jsonb` column (PostgreSQL), or other backends
-through a JSON representation of the history encapsulated by
-[LLM::Session](https://0x1eef.github.io/x/llm.rb/LLM/Session.html)
-&ndash; inclusive of tool metadata as well:
+llm.rb uses Ruby's `Monitor` class to ensure thread safety at the provider
+level, allowing you to share a single provider instance across multiple threads
+while maintaining state isolation through thread-local contexts. This design
+enables efficient resource sharing while preventing race conditions in
+concurrent applications:
 
-* Process 1
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
+# Thread-safe providers - create once, use everywhere
 llm = LLM.openai(key: ENV["KEY"])
-ses = LLM::Session.new(llm)
-ses.talk "Howdy partner"
-ses.talk "I'll see you later"
-ses.save(path: "session.json")
-```
-* Process 2
-```ruby
-#!/usr/bin/env ruby
-require "llm"
-require "pp"
 
-llm = LLM.openai(key: ENV["KEY"])
-ses = LLM::Session.new(llm)
-ses.restore(path: "session.json")
-ses.talk "Howdy partner. I'm back"
-pp ses.messages
-```
-
-But how does it work without a file ? The [LLM::Session](https://0x1eef.github.io/x/llm.rb/LLM/Session.html)
-class implements `#to_json` and it can be used to obtain a JSON representation
-of a session that can be stored in a `jsonb` column in PostgreSQL, or any
-other storage backend. The session can then be restored from the JSON
-representation via the restore method and its `string` argument:
-
-```ruby
-#!/usr/bin/env ruby
-require "llm"
+# Each thread should have its own context for state isolation
+Thread.new do
+  ctx = LLM::Context.new(llm)  # Thread-local context
+  ctx.talk("Hello from thread 1")
+end
 
-llm = LLM.openai(key: ENV["KEY"])
-ses1 = LLM::Session.new(llm)
-ses1.talk "Howdy partner"
-ses1.talk "I'll see you later"
-
-json = ses1.to_json
-ses2 = LLM::Session.new(llm)
-ses2.restore(string: json)
-ses2.talk "Howdy partner. I'm back"
+Thread.new do
+  ctx = LLM::Context.new(llm)  # Thread-local context
+  ctx.talk("Hello from thread 2")
+end
 ```
 
-### Tools
+#### Performance Tuning
 
-#### LLM::Function
-
-The following example demonstrates [LLM::Function](https://0x1eef.github.io/x/llm.rb/LLM/Function.html)
-and how it can define a local function (which happens to be a tool), and how
-a provider (such as OpenAI) can then detect when we should call the function.
-Its most notable feature is that it can act as a closure and has access to
-its surrounding scope, which can be useful in some situations:
+llm.rb's JSON adapter system lets you swap JSON libraries for better
+performance in high-throughput applications. The library supports stdlib JSON,
+Oj, and Yajl, with Oj typically offering the best performance. Additionally,
+you can enable HTTP connection pooling using the optional `net-http-persistent`
+gem to reduce connection overhead in production environments:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
-llm  = LLM.openai(key: ENV["KEY"])
-tool = LLM.function(:system) do |fn|
-  fn.description "Run a shell command"
-  fn.params do |schema|
-    schema.object(command: schema.string.required)
-  end
-  fn.define do |command:|
-    ro, wo = IO.pipe
-    re, we = IO.pipe
-    Process.wait Process.spawn(command, out: wo, err: we)
-    [wo,we].each(&:close)
-    {stderr: re.read, stdout: ro.read}
-  end
-end
-
-ses = LLM::Session.new(llm, tools: [tool])
-ses.talk "Your task is to run shell commands via a tool.", role: :user
-
-ses.talk "What is the current date?", role: :user
-ses.talk ses.functions.map(&:call) # report return value to the LLM
+# Swap JSON libraries for better performance
+LLM.json = :oj  # Use Oj for faster JSON parsing
 
-ses.talk "What operating system am I running?", role: :user
-ses.talk ses.functions.map(&:call) # report return value to the LLM
-
-##
-# {stderr: "", stdout: "Thu May  1 10:01:02 UTC 2025"}
-# {stderr: "", stdout: "FreeBSD"}
+# Enable HTTP connection pooling for high-throughput applications
+llm = LLM.openai(key: ENV["KEY"]).persist!  # Uses net-http-persistent when available
 ```
 
-#### LLM::Tool
-
-The [LLM::Tool](https://0x1eef.github.io/x/llm.rb/LLM/Tool.html) class can be used
-to implement a [LLM::Function](https://0x1eef.github.io/x/llm.rb/LLM/Function.html)
-as a class. Under the hood, a subclass of [LLM::Tool](https://0x1eef.github.io/x/llm.rb/LLM/Tool.html)
-wraps an instance of [LLM::Function](https://0x1eef.github.io/x/llm.rb/LLM/Function.html)
-and delegates to it.
+#### Model Registry
 
-The choice between [LLM::Function](https://0x1eef.github.io/x/llm.rb/LLM/Function.html)
-and [LLM::Tool](https://0x1eef.github.io/x/llm.rb/LLM/Tool.html) is often a matter of
-preference but each carry their own benefits. For example, [LLM::Function](https://0x1eef.github.io/x/llm.rb/LLM/Function.html)
-has the benefit of being a closure that has access to its surrounding context and
-sometimes that is useful:
+llm.rb includes a local model registry that provides metadata about model
+capabilities, pricing, and limits without requiring API calls. The registry is
+shipped with the gem and sourced from https://models.dev, giving you access to
+up-to-date information about context windows, token costs, and supported
+modalities for each provider:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
-class System < LLM::Tool
-  name "system"
-  description "Run a shell command"
-  param :command, String, "The command to execute", required: true
-
-  def call(command:)
-    ro, wo = IO.pipe
-    re, we = IO.pipe
-    Process.wait Process.spawn(command, out: wo, err: we)
-    [wo,we].each(&:close)
-    {stderr: re.read, stdout: ro.read}
-  end
-end
-
-llm = LLM.openai(key: ENV["KEY"])
-ses = LLM::Session.new(llm, tools: [System])
-ses.talk "Your task is to run shell commands via a tool.", role: :user
-
-ses.talk "What is the current date?", role: :user
-ses.talk ses.functions.map(&:call) # report return value to the LLM
-
-ses.talk "What operating system am I running?", role: :user
-ses.talk ses.functions.map(&:call) # report return value to the LLM
-
-##
-# {stderr: "", stdout: "Thu May  1 10:01:02 UTC 2025"}
-# {stderr: "", stdout: "FreeBSD"}
+# Access model metadata, capabilities, and pricing
+registry = LLM.registry_for(:openai)
+model_info = registry.limit(model: "gpt-4.1")
+puts "Context window: #{model_info.context} tokens"
+puts "Cost: $#{model_info.cost.input}/1M input tokens"
 ```
 
-### Files
+## More Examples
 
-#### Create
+#### Responses API
 
-The OpenAI and Gemini providers provide a Files API where a client can upload files
-that can be referenced from a prompt, and with other APIs as well. The following
-example uses the OpenAI provider to describe the contents of a PDF file after
-it has been uploaded. The file (a specialized instance of
-[LLM::Response](https://0x1eef.github.io/x/llm.rb/LLM/Response.html)
-) is given as part of a prompt that is understood by llm.rb:
+llm.rb also supports OpenAI's Responses API through `llm.responses` and
+`ctx.respond`. This API can maintain response state server-side and can reduce
+how much conversation state needs to be sent on each turn:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
-require "pp"
 
 llm = LLM.openai(key: ENV["KEY"])
-ses = LLM::Session.new(llm)
-file = llm.files.create(file: "/tmp/llm-book.pdf")
-res = ses.talk ["Tell me about this file", file]
-pp res.content
-```
-
-### Prompts
+ctx = LLM::Context.new(llm)
 
-#### Multimodal
+ctx.respond("Your task is to answer the user's questions", role: :developer)
+res = ctx.respond("What is the capital of France?")
+puts res.output_text
+```
 
-LLMs are great with text, but many can also handle images, audio, video,
-and URLs. With llm.rb you pass those inputs by tagging them with one of
-the following methods. And for multipart prompts, we can pass an array
-where each element is a part of the input. See the example below for
-details, in the meantime here are the methods to know for multimodal
-inputs:
+#### Context Persistence
 
-* `ses.image_url` for an image URL
-* `ses.local_file` for a local file
-* `ses.remote_file` for a file already uploaded via the provider's Files API
+Contexts can be serialized and restored across process boundaries. This makes
+it possible to persist conversation state in a file, database, or queue and
+resume work later:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
 llm = LLM.openai(key: ENV["KEY"])
-ses = LLM::Session.new(llm)
-res = ses.talk ["Tell me about this image URL", ses.image_url(url)]
-res = ses.talk ["Tell me about this PDF", ses.remote_file(file)]
-res = ses.talk ["Tell me about this image", ses.local_file(path)]
+ctx = LLM::Context.new(llm)
+ctx.talk("Hello")
+ctx.talk("Remember that my favorite language is Ruby")
+ctx.save(path: "context.json")
+
+restored = LLM::Context.new(llm)
+restored.restore(path: "context.json")
+res = restored.talk("What is my favorite language?")
+puts res.content
 ```
 
-### Audio
-
-#### Speech
+#### Agents
 
-Some but not all providers implement audio generation capabilities that
-can create speech from text, transcribe audio to text, or translate
-audio to text (usually English). The following example uses the OpenAI provider
-to create an audio file from a text prompt. The audio is then moved to
-`${HOME}/hello.mp3` as the final step:
+Agents in llm.rb are reusable, preconfigured assistants that automatically
+execute tool calls and maintain conversation state. Unlike contexts which
+require manual tool execution, agents automatically handle the tool call loop,
+making them ideal for autonomous workflows where you want the LLM to
+independently use available tools to accomplish tasks:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
+class SystemAdmin < LLM::Agent
+  model "gpt-4.1"
+  instructions "You are a Linux system admin"
+  tools Shell
+  schema Result
+end
+
 llm = LLM.openai(key: ENV["KEY"])
-res = llm.audio.create_speech(input: "Hello world")
-IO.copy_stream res.audio, File.join(Dir.home, "hello.mp3")
+agent = SystemAdmin.new(llm)
+res = agent.talk("Run 'date'")
 ```
 
-#### Transcribe
+#### Cost Tracking
 
-The following example transcribes an audio file to text. The audio file
-(`${HOME}/hello.mp3`) was theoretically created in the previous example,
-and the result is printed to the console. The example uses the OpenAI
-provider to transcribe the audio file:
+llm.rb provides built-in cost estimation that works without making additional
+API calls. The cost tracking system uses the local model registry to calculate
+estimated costs based on token usage, giving you visibility into spending
+before bills arrive. This is particularly useful for monitoring usage in
+production applications and setting budget alerts:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
 llm = LLM.openai(key: ENV["KEY"])
-res = llm.audio.create_transcription(
-  file: File.join(Dir.home, "hello.mp3")
-)
-puts res.text # => "Hello world."
+ctx = LLM::Context.new(llm)
+ctx.talk "Hello"
+puts "Estimated cost so far: $#{ctx.cost}"
+ctx.talk "Tell me a joke"
+puts "Estimated cost so far: $#{ctx.cost}"
 ```
 
-#### Translate
+#### Multimodal Prompts
 
-The following example translates an audio file to text. In this example
-the audio file (`${HOME}/bomdia.mp3`) is theoretically in Portuguese,
-and it is translated to English. The example uses the OpenAI provider,
-and at the time of writing, it can only translate to English:
+Contexts provide helpers for composing multimodal prompts from URLs, local
+files, and provider-managed remote files. These tagged objects let providers
+adapt the input into the format they expect:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
 
 llm = LLM.openai(key: ENV["KEY"])
-res = llm.audio.create_translation(
-  file: File.join(Dir.home, "bomdia.mp3")
-)
-puts res.text # => "Good morning."
-```
-
-### Images
+ctx = LLM::Context.new(llm)
 
-#### Create
-
-Some but not all LLM providers implement image generation capabilities that
-can create new images from a prompt, or edit an existing image with a
-prompt. The following example uses the OpenAI provider to create an
-image of a dog on a rocket to the moon. The image is then written to
-`${HOME}/dogonrocket.png` as the final step:
-
-```ruby
-#!/usr/bin/env ruby
-require "llm"
-llm = LLM.openai(key: ENV["KEY"])
-res = llm.images.create(prompt: "a dog on a rocket to the moon")
-IO.copy_stream res.images[0], File.join(Dir.home, "dogonrocket.png")
+res = ctx.talk ["Describe this image", ctx.image_url("https://example.com/cat.jpg")]
+puts res.content
 ```
 
-#### Edit
+#### Audio Generation
 
-The following example is focused on editing a local image with the aid
-of a prompt. The image (`/tmp/llm-logo.png`) is returned to us with a hat.
-The image is then written to `${HOME}/logo-with-hat.png` as
-the final step:
+llm.rb supports OpenAI's audio API for text-to-speech generation, allowing you
+to create speech from text with configurable voices and output formats. The
+audio API returns binary audio data that can be streamed directly to files or
+other IO objects, enabling integration with multimedia applications:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
+
 llm = LLM.openai(key: ENV["KEY"])
-res = llm.images.edit(
-  image: "/tmp/llm-logo.png",
-  prompt: "add a hat to the logo",
-)
-IO.copy_stream res.images[0], File.join(Dir.home, "logo-with-hat.png")
+res = llm.audio.create_speech(input: "Hello world")
+IO.copy_stream res.audio, File.join(Dir.home, "hello.mp3")
 ```
 
-#### Variations
+#### Image Generation
 
-The following example is focused on creating variations of a local image.
-The image (`/tmp/llm-logo.png`) is returned to us with five different variations.
-The images are then written to `${HOME}/logo-variation0.png`, `${HOME}/logo-variation1.png`
-and so on as the final step:
+llm.rb provides access to OpenAI's DALL-E image generation API through a
+unified interface. The API supports multiple response formats including
+base64-encoded images and temporary URLs, with automatic handling of binary
+data streaming for efficient file operations:
 
 ```ruby
 #!/usr/bin/env ruby
 require "llm"
+
 llm = LLM.openai(key: ENV["KEY"])
-res = llm.images.create_variation(
-  image: "/tmp/llm-logo.png",
-  n: 5
-)
-res.images.each.with_index do |image, index|
-  IO.copy_stream image,
-                 File.join(Dir.home, "logo-variation#{index}.png")
-end
+res = llm.images.create(prompt: "a dog on a rocket to the moon")
+IO.copy_stream res.images[0], File.join(Dir.home, "dogonrocket.png")
 ```
 
-### Embeddings
-
-#### Text
+#### Embeddings
 
-The
-[`LLM::Provider#embed`](https://0x1eef.github.io/x/llm.rb/LLM/Provider.html#embed-instance_method)
-method returns vector embeddings for one or more text inputs. A common
-use is semantic search (store vectors, then query for similar text):
+llm.rb's embedding API generates vector representations of text for semantic
+search and retrieval-augmented generation (RAG) workflows. The API supports
+batch processing of multiple inputs and returns normalized vectors suitable for
+vector similarity operations, with consistent dimensionality across providers:
 
 ```ruby
 #!/usr/bin/env ruby
@@ -760,52 +553,32 @@ puts res.class
 puts res.embeddings.size
 puts res.embeddings[0].size
 
-##
 # LLM::Response
 # 3
 # 1536
 ```
 
-### Models
+## Real-World Example: Relay
 
-#### List
+See how these pieces come together in a complete application architecture with
+[Relay](https://github.com/llmrb/relay), a production-ready LLM application
+built on llm.rb that demonstrates:
 
-Almost all LLM providers provide a models endpoint that allows a client to
-query the list of models that are available to use. The list is dynamic,
-maintained by LLM providers, and it is independent of a specific llm.rb
-release:
+- Context management across requests
+- Tool composition and execution
+- Concurrent workflows
+- Cost tracking and observability
+- Production deployment patterns
 
-```ruby
-#!/usr/bin/env ruby
-require "llm"
-require "pp"
-
-##
-# List all models
-llm = LLM.openai(key: ENV["KEY"])
-llm.models.all.each do |model|
-  puts "model: #{model.id}"
-end
-
-##
-# Select a model
-model = llm.models.all.find { |m| m.id == "gpt-3.5-turbo" }
-ses = LLM::Session.new(llm, model: model.id)
-res = ses.talk "Hello #{model.id} :)"
-pp res.content
-```
+Watch the screencast:
 
-## Install
+[![Watch the llm.rb screencast](https://img.youtube.com/vi/Jb7LNUYlCf4/maxresdefault.jpg)](https://www.youtube.com/watch?v=x1K4wMeO_QA)
 
-llm.rb can be installed via rubygems.org:
+## Installation
 
-	gem install llm.rb
-
-## Sources
-
-* [GitHub.com](https://github.com/llmrb/llm.rb)
-* [GitLab.com](https://gitlab.com/llmrb/llm.rb)
-* [Codeberg.org](https://codeberg.org/llmrb/llm.rb)
+```bash
+gem install llm.rb
+```
 
 ## License