AdalFlow Repo Skill

Use this skill when a task involves the AdalFlow Python package: building model-agnostic LLM task pipelines, structured outputs, RAG, agents, tool calling, evaluation, prompt/few-shot/text-gradient optimization, or tracing.

AdalFlow uses PyTorch-like Component building blocks. A typical workflow combines Prompt, ModelClient, Generator, DataClass/parser objects, retrievers, tools, agents, metrics, and tracing callbacks.

First checks

1. Read repository provenance before assuming this skill is current for a checkout.
2. For a package install/import task, read package overview and root troubleshooting.
3. For a quick environment smoke, run check_adalflow_environment.py in the target Python environment.
4. Then route to the most specific sub-skill below.

Install/import guidance

For AdalFlow 1.1.1, use Python 3.9+ and include the OpenAI SDK for reliable top-level import because the current Generator module imports OpenAI response event types:

python -m pip install "adalflow[openai]"
python - <<'PY'
import adalflow as adal
print(adal.__version__)
from adalflow import Component, DataClass, Generator, Agent, Runner
print("adalflow import ok")
PY

Install additional extras only for workflows that need them, such as provider SDKs, FAISS/LanceDB/Postgres/Qdrant retrieval, MCP, torch/transformers, datasets, or SQLAlchemy. Do not treat a missing API key, service, dataset, or accelerator as a base package failure.

Route map

Core components and structured I/O

Read core-components-and-structured-io for:

• Component, DataComponent, Sequential, ComponentList, Prompt, DataClass, and required_field.
• JSON/YAML/list/int/float/bool parsers and DataClassParser structured-output prompts.
• Service-free mini-pipeline checks that do not call a model provider.

Model clients, generators, and embedders

Read model-client-and-generator-workflows for:

• ModelClient, Generator, GeneratorOutput, Embedder, and BatchEmbedder.
• prompt_kwargs, model_kwargs, ModelType, output processors, cache behavior, and async/streaming basics.
• Provider extras and no-credential fake-client tests.

Retrieval, RAG, and data pipelines

Read retrieval-rag-and-data-pipelines for:

• Document, TextSplitter, LocalDB, ToEmbeddings, and RetrieverOutputToContextStr.
• BM25Retriever, FAISSRetriever, LanceDB, Postgres, Qdrant, and RAG context assembly.
• Optional vector-store, embedding-dimension, stale-index, and service troubleshooting.

Agents, tools, streaming, HITL, and MCP

Read agents-tools-and-streaming for:

• FunctionTool, ToolManager, FunctionDefinition, FunctionOutput, and ToolOutput.
• Agent, Runner, ReActAgent, RunnerResult, StepOutput, and stream events.
• Human approval/permission handlers, safe tool patterns, and optional MCP tools.

Evaluation and optimization

Read evaluation-and-optimization for:

• AnswerMatchAcc, RetrieverEvaluator, cautious LLMasJudge, and dataset loaders.
• Parameter, GradComponent, AdalComponent, Trainer, BootstrapFewShot, TGDOptimizer, EvalFnToTextLoss, LLMAsTextLoss, and optimize_anything.
• Training/benchmark workflows that require provider credentials, data, and runtime budget.

Tracing, observability, and configuration

Read tracing-observability-and-configuration for:

• setup_env, logging helpers, config loading, and default artifact roots.
• GeneratorStateLogger, GeneratorCallLogger, callbacks, spans, trace providers, and optional MLflow.
• Debug artifacts, path hygiene, and trace/log troubleshooting.

Cross-workflow rules

• Start service-free when possible: validate DataClass, parser, TextSplitter, metrics, fake ModelClient, or fake Runner behavior before adding providers.
• Choose provider extras only after the required ModelClient is known. API credentials and network calls are workflow prerequisites, not skill smoke checks.
• For RAG, separate data preparation, retrieval index construction, context formatting, generator configuration, evaluation metrics, and tracing; each has a different owner sub-skill.
• For agents, verify tools independently with FunctionTool/ToolManager before letting an LLM planner call them.
• For optimization, define evaluation and loss contracts before running Trainer.fit; prefer Trainer.diagnose or metric smokes before expensive training.
• For tracing/logging, keep artifacts in explicit writable directories and avoid logging secrets, provider payloads, or machine-specific paths.

When not to use this skill

• The task is only about editing this repository's release process, CI, or publishing scripts.
• The task needs a different package's APIs and only mentions AdalFlow as background context.
• The user asks Creator to perform downstream research directly; switch to Researcher mode instead of using this operating graph here.