Examples

Runnable examples live under examples/ — one per feature. Most require local Ollama (no API keys); check each directory's README for exact commands.

Example	What it shows
basic_pipeline	Minimal YAML pipeline, no API keys
branching	Conditional edges + Flow API
parallel	Concurrent branches + typed State reducers
map_repair_plans	Dynamic fan-out (Map) + {key} prompt templates + typed State
human_in_loop	Approve/Edit LLM output via Interrupt + loop() + resume (Python and YAML)
ask_strategies	Validate interrupt answers with Ask — regex (capture a promo code), equals, and an LLM model classifier (offline, no API key)
react_agent	ReAct agent loop with a calculator tool and live token streaming
memory_assistant	Long-term memory: LLM fact extraction, MemoryStore + provider-aware embedder, context injection
memory_chat	Multi-user streaming chat — owner picked at the console, per-owner memory (${owner}), live tokens, auto fact extraction
harness_agent	flow.harness() — parallel tool calls in one round + __error__ fallback
hello_workflow	The same deterministic workflow at three levels — YAML, Flow DSL, low-level graph; no API key, CLI-runnable
hello_llm	Minimal LLM workflow (llm_chat + transform) — CLI-runnable with durable SQLite checkpoints
poem_chat	Two-agent poem chat with human approval — context_builder → llm_chat poet → critic → interrupt; the approval answer is classified by a small LLM (no hard-coded keywords) and a failed approval loops the poem back for a rewrite (teff chat)
agent_approval	Tool approval (HITL) — every tool call pauses for human sign-off and resumes
agent_resilience	Retries, model failover, context trimming and a token budget (mocked, no API key)
skills	Skills folder (SKILL.md) — instructions + tool scoping on a harness agent
pdf_agent	Skill with its own tools — vendored pdf skill whose bundled scripts run via the shell tool
mcp	ReAct agent calling tools from an MCP server (stdio)
plugins	Custom nodes/tools via decorators and via subclasses; offline + agent variants
streaming	Live LLM tokens + graph events via graph.stream()
observability	langfuse-style trace viewer — GraphObserver captures topology, node spans and full LLM prompt/response into SQLite; FastAPI dashboard UI + push exporters (webhook/langfuse/langsmith) and teff obs-server
structured_output	Schema-validated LLM JSON via output_type / json_schema
rag_search	RAG over a local CSV, in-memory store
rag_stores	Same RAG agent on every vector store
checkpoint_resume	Crash mid-LLM run, resume from the failed node — no tokens wasted
checkpoint_stores	Durable workflow on file/sqlite/pg
release_features	Release API tour — validation, typed errors, teff eval, cost reports, response cache (mocked, no API key)
simple_router	Minimal Flow.route() supervisor — two agents, a bounded loop (can't hang), offline tests
command_routing	Dynamic per-node routing with Command — update+goto, goto=Command.STOP, edge bypass (offline, no API key)
yaml_compose	Pure-YAML composition — include:, loop, command routing, new transform actions (offline, no API key)
fraud_gate	Production FastAPI payment gate — LLM scorer + Command routing (approve / mid-risk human review / deny-and-stop)
service_desk	Default supervisor() chat router — one-word dispatch, done_keys/fallback_agent guards, bounded loop and a human Interrupt deploy gate (Russian support desk)
repair-ai-chat	Full FastAPI app — one ReAct coordinator driving specialist tools (extract, plan, materials, estimate, QA), RAG, streaming (Russian repair workflow)
channels	One durable Assistant over every transport — HTTP/SSE (teff serve), Telegram (teff bot) and terminal (teff chat); zero-code channels: YAML block
channels/supervisor	Multi-agent supervisor (planner/coder/QA) wrapped in the channels: block — llm_chat JSON verdicts, Command routing, loop-until-pass refine, human approve gate (Russian prompts)
channels/rag_ingest	Grow the vector store from any channel — llm_chat normalizes a raw row, then the rag_ingest tool chunks/embeds it to SQLite; query the grown base with rag

All LLM examples use llama3.1:8b (ollama pull llama3.1:8b); pdf_agent uses qwen2.5:7b (ollama pull qwen2.5:7b).