Examples in this section
| Example | What it shows |
|---|---|
| Data extraction pipeline | Use @generative with a typed return to pull structured data from unstructured text |
| Legacy code integration | Apply @mify to existing Python classes so the model can act on them |
| Resilient RAG with fallback | Build a FAISS retrieval pipeline with an LLM relevance filter before generation |
| Traced generation loop | Enable OpenTelemetry application and backend traces with two environment variables |
All example categories
The repository contains many more runnable examples than the four documented above. Every category has its ownREADME.md and one or more .py files ready
to run.
Core concepts
| Category | What it shows |
|---|---|
instruct_validate_repair/ | The IVR loop end-to-end: basic generation, adding requirements, automatic repair on failure, custom validators |
generative_slots/ | @generative functions with typed returns, pipeline composition, ChatContext persona injection, pre/postcondition checks |
context/ | Context inspection, sampling with context trees, parallel context branches |
sessions/ | Custom session types and backend selection |
Data and documents
| Category | What it shows |
|---|---|
information_extraction/ | Named entity recognition and type-safe structured extraction with Pydantic |
mobject/ | Table queries and transformations using MObject structured data types |
mify/ | @mify on existing classes — custom string representations, field filtering, funcs_include |
rag/ | FAISS vector search, @generative bool relevance filter, grounding_context for grounded generation |
Agents and tools
| Category | What it shows |
|---|---|
agents/ | ReACT reasoning-and-acting loop, multi-turn tool workflows |
tools/ | @tool definition, code interpreter integration, tool argument validation, safe eval patterns |
mini_researcher/ | Complete research assistant: multi-model architecture, document retrieval, safety checks, custom validation pipeline |
Safety and validation
| Category | What it shows |
|---|---|
safety/ | GuardianCheck for harm, jailbreak, profanity, social bias, violence, and groundedness; shared backend pattern |
Integration and deployment
| Category | What it shows |
|---|---|
m_serve/ | Deploying Mellea programs as REST APIs with production deployment patterns |
library_interop/ | LangChain message conversion, OpenAI format compatibility, cross-library workflows |
mcp/ | MCP tool creation, Claude Desktop integration, Langflow integration |
bedrock/ | Amazon Bedrock backend configuration and usage |
Performance and advanced sampling
| Category | What it shows |
|---|---|
aLora/ | Training aLoRA adapters for fast constraint checking; performance optimisation |
intrinsics/ | Answer relevance, hallucination detection, citation validation, context relevance — specialised adapter-backed checks |
sofai/ | Two-tier sampling: fast-model iteration with escalation to a slow model; cost optimisation |
Multimodal
| Category | What it shows |
|---|---|
image_text_models/ | Vision-language models, ImageBlock, multimodal prompting, backend support matrix |
Observability
| Category | What it shows |
|---|---|
telemetry/ | OpenTelemetry application and backend traces; span export configuration |
Experimental
| Category | What it shows |
|---|---|
melp/ | ⚠️ Experimental lazy evaluation — thunks, deferred execution, advanced control flow |
Running the examples
All examples are in thedocs/examples/ directory of the repository. Unless
otherwise noted, run them with:
/// script block and can be
run with uv run instead:
start_session() with no arguments connects to a local
Ollama instance running IBM Granite 4 Micro
(granite4:micro). Make sure Ollama is running before you execute any example.