emergentflow 0.2.2

Emergent Flow — core SDK and graph intermediate representation (IR).

Emergent Flow

The Visual Architecture Platform for Data Science, Analytics & Machine Learning

Emergent Flow is an infinite-canvas visual development platform that unifies the fragmented data ecosystem. It bridges the divide between low-code/no-code visual tools (which lack flexibility and lock users into proprietary environments) and code-first environments (which are linear, complex, and prone to technical friction).

Through a Figma-like multiplayer whiteboard interface, Emergent Flow lets you visually architect entire data lifecycles — from raw ETL and classical hypothesis testing, to AutoML, custom deep-learning architectures, and generative-AI multi-agent orchestration. Crucially, the visual canvas maps 1:1 with an underlying, optimized open-source Python SDK: every drag-and-drop connection compiles to beautiful, production-grade, human-readable Python — eliminating vendor lock-in and preserving true developer freedom.

Status: Phases 1–2 are implemented and CI-tested; Phase 3 is partially built. Done: the graph IR with its node contract/registry and type-validated connections (Epics 1, 3); the ef.compile_to_code / ef.execute codegen engine, including the declarative nn.Module seam (Epic 2); a broad node library spanning data/clean/stats/viz, dedicated per-family statistical modeling nodes (linear regression, GLM, GAM, mixed models, Bayesian), feature transformation nodes (scaling, encoding, discretization, feature generation), a full scikit-learn adapter (Epic 8), SHAP-based model explainability with diagnostic plots (ADR 0020), custom-code nodes for user-defined Python transforms, data-warehouse connectors — DuckDB, Postgres, BigQuery, Redshift (Epic 13) — and LLM nodes behind an injected, replayable client seam plus a Prompt Lab for eval/labeling (Epic 9); the ui/ canvas (React Flow + Zustand) with CodeMirror 6 code editing, column-aware parameter widgets, expandable inspector panels, and markdown notes for graph annotation; a bundled local FastAPI server (emergentflow serve) with streaming execution, DAG-aware incremental caching, and visual results (Epics 5–7); a collaboration layer (emergentflow/collab/) with an in-app agent chat panel and multi-backend agent adapters (Claude, Gemini, Codex, OpenCode) letting an AI agent propose graph edits as reviewable GraphMutations alongside a human on the same canvas (Epic 14); and a unified connection manager for warehouse, LLM, and agent profiles. Still stubs: full multi-step agentic orchestration / LangGraph codegen (Epic 10), RAG (Epic 11), and real-time multiplayer. See epics/README.md for the epic-by-epic delivery status, the Technical Roadmap for the engineering decomposition, and the Product Proposal for the vision.

---

Installation

pip install emergentflow[server]  # SDK + server + bundled canvas
emergentflow serve                # opens localhost:8765 in your browser

A bare pip install emergentflow installs the SDK for programmatic use (notebooks, scripts, CI). The [server] extra adds FastAPI/Uvicorn so emergentflow serve (alias emergentflow lab) can boot the local canvas. Optional extras for specific node families:

pip install emergentflow[all]       # everything — server + all optional extras
pip install emergentflow[llm]       # LiteLLM gateway for LLM nodes
pip install emergentflow[explain]   # SHAP-based model explainability
pip install emergentflow[bayes]     # Bayesian modeling (PyMC/Bambi/ArviZ)
pip install emergentflow[bigquery]  # BigQuery warehouse adapter
pip install emergentflow[postgres]  # Postgres warehouse adapter
pip install emergentflow[redshift]  # Redshift warehouse adapter
pip install emergentflow[mcp]       # MCP tool wrapper for agent collaboration

Combine extras as needed: pip install emergentflow[server,llm,explain], or use pip install emergentflow[all] to get everything in one shot.

---

Why Emergent Flow

Traditional data science happens in linear, isolated environments like Jupyter notebooks, which obscure data lineage, make collaboration difficult, and hide the structure of complex pipelines. Emergent Flow brings the design paradigm of Figma and Miro to the DE / DS / ML / GenAI stack:

• Complete lifecycle visibility — trace data lineage transparently from a raw SQL database, through feature transformations, and into deep-learning tensors or LLM prompts.
• Glass-box code generation — clicking any node or pipeline path reveals clean, PEP8-compliant, runnable Python that can be exported straight to Git.
• Frictionless AI evolution — the architecture grows naturally from structured data processing to deep-learning execution and multi-agent generative-AI routing, all on the same canvas.

[ Data Source ] ──> [ Data Prep / Impute ] ──> [ Custom DL Layer ] ──> [ Agent Orchestrator ]
       │                                              │                        │
       └──> [ Classical Statistics ]                  └──> [ Evaluator ]       └──> [ API / App ]

---

Architecture

Emergent Flow ships as one repository and one bundled pip install emergentflow[server] — the JupyterLab model: the Python data-science engine and a local web canvas in a single install, launched with emergentflow serve (alias emergentflow lab). A bare pip install emergentflow gives the SDK for programmatic use; the [server] extra adds FastAPI/Uvicorn for the canvas. The canvas and the SDK are separate toolchains that couple only through published contract artifacts (the IR JSON Schema, the node catalog, the GraphMutation/session-event schemas, and the connection-validation rules-as-data) — the UI never imports Python. See ADR 0013.

    pip install emergentflow[server]  →  emergentflow serve  (alias: emergentflow lab)
┌──────────────────────────────────────────────────────────┐
│   ui/  —  CANVAS (React Flow / @xyflow, Zustand, Vite)    │  bundled into the wheel
│         talks only via IR schema · codegen · rules-data  │  (never imports emergentflow)
│         · catalog · mutation/session-event schemas        │
└────────────────────────────┬─────────────────────────────┘
                             │ localhost REST + SSE  (no shared import)
                             ▼
┌──────────────────────────────────────────────────────────┐
│   emergentflow/server/  —  LOCAL FastAPI/Uvicorn server  │  calls ef.* in-process
│         streaming execute · DAG cache · sessions/collab   │  (no Celery / Redis / sandbox)
└────────────────────────────┬─────────────────────────────┘
                             │ (pure functions over one IR)
                             ▼
┌──────────────────────────────────────────────────────────┐
│   emergentflow/  —  CORE PYTHON SDK & EXECUTION          │
│  Pandas | Statsmodels | Scikit-Learn | Plotly | DuckDB    │
│  SHAP | SQLAlchemy/BigQuery/Redshift | LiteLLM | PyTorch  │
└──────────────────────────────────────────────────────────┘

The happy path is local and in-process (§A6 of the roadmap). The bundled package is a single-user, local-first app — no Celery, no Redis, no sandbox, no multi-tenancy. The enterprise scale-out (distributed/sandboxed execution, Redis + object-store caching, auth, multiplayer) is deferred to a future gated hosted product (the dbt-Cloud to this dbt-core), not the bundled install.

Core design principles

These cross-cutting decisions shape nearly every part of the system. Each is recorded as a formal Architecture Decision Record in docs/adr/ and derives from §A of the roadmap:

1. The graph is the single source of truth; code is a compiled artifact. The serialized graph (the intermediate representation, or IR) is canonical. Python is a one-way build output, not a co-equal representation requiring bidirectional sync.
2. Execute the IR, not the generated string. Production interprets the IR directly by calling SDK functions; the generated Python is for display and export only. Equivalence between compile_to_code(ir) and execute(ir) is enforced as a hard invariant in CI.
3. The SDK supports two paradigms. A functional pipeline (DE / stats / classical ML / reporting) and a declarative module/graph definition (PyTorch architectures, LangGraph agent graphs) are first-class from day one.
4. Storage is tiered (hosted tier). On the bundled happy path the cache is a simple in-memory + on-disk store. The tiered design — Redis for cache metadata/hashes/small results, a disk/object store for large artifacts via Arrow / Parquet / safetensors — is a hosted-product concern (see ADR 0013 and §A6 of the roadmap), not something the local install needs.
5. Non-deterministic, credentialed I/O is an injected client, never inline. LLM calls (ADR 0017) and data-warehouse queries (ADR 0018) both route through a client object execute/the compiled entry point receives as a parameter — never a call made directly inside codegen/execute — so a ReplayClient keeps the ADR-0002 equivalence gate value-exact and offline, and the IR itself stays free of live connections or credentials.
6. Collaboration state lives beside the graph, never on it (ADR 0019). Agent proposals, review threads, and gates are separate models in emergentflow/collab/; the Graph IR schema is untouched, so the product works identically with or without agents attached.

See the Architecture Decision Records for the full context, decision, and consequences of each: ADR 0001, ADR 0002, ADR 0003, ADR 0004, ADR 0005, ADR 0006, ADR 0007, ADR 0008, ADR 0009, ADR 0010, ADR 0011, ADR 0012, ADR 0013, ADR 0014, ADR 0015, ADR 0016, ADR 0017, ADR 0018, ADR 0019, ADR 0020.

Example of generated code

The visual builder compiles structured SDK objects rather than arbitrary code blocks, so the output stays clean and maintainable:

import emergentflow as ef

# 1. Data ingestion & imputation
df = ef.data.load_csv("customer_churn.csv")
df_clean = ef.clean.impute_missing(df, columns=["age", "income"], strategy="median")

# 2. Statistical validation
stat_results = ef.stats.anova(df_clean, dv="churn_risk", between="segment")

# 3. Model architecture & pipeline training
model, metrics = ef.ml.train_classifier(
    data=df_clean,
    target="churn_risk",
    model_type="random_forest",
    optimize_for="f1",
)

# 4. Export artifacts
ef.reports.generate_html_summary(model, metrics, output_path="churn_report.html")

The same ef.compile_to_code entry point dispatches on paradigm: declarative nn.Module graphs compile through a libcst-based generator to idiomatic PyTorch class definitions (an __init__ of layers plus a forward chain) rather than a flat script. See the Declarative Codegen Seam for the worked example.

---

Technology Stack

Layer	Technologies
Canvas (ui/ tree, bundled)	React + @xyflow/react (React Flow), Zustand, CodeMirror 6, Vite, TypeScript
Local server (emergentflow/server/)	FastAPI + Uvicorn (optional [server] extra) — REST, SSE streaming execution, session/collab routes, connection management, in-process ef.*
Collaboration (emergentflow/collab/)	GraphMutation proposals, sessions, review threads, gates, personas, in-app agent chat, multi-backend agent adapters (Claude, Gemini, Codex, OpenCode); optional MCP tool wrapper ([mcp] extra, FastMCP) — ADR 0019
Backend engine (hosted tier)	Celery, Redis — deferred to the hosted product (ADR 0013 / §A6)
Data wrangling	Pandas, PyArrow (Parquet)
Feature transforms (emergentflow/nodes/examples/)	Scaling, categorical encoding, discretization, polynomial/interaction feature generation — all via scikit-learn
Data warehouses (emergentflow/data/warehouse/)	DuckDB (in-process, credential-free), Postgres (SQLAlchemy + psycopg, [postgres]), BigQuery ([bigquery]), Redshift ([redshift]); sqlglot for dialect SQL — ADR 0018
Statistics	Statsmodels, SciPy; dedicated per-family nodes (linear regression, GLM, GAM, mixed models); optional Bayesian family via PyMC/Bambi/ArviZ ([bayes])
Machine learning	Scikit-Learn
Model explainability (emergentflow/explain/)	SHAP-based feature attribution, error tables, diagnostic plots (residuals, calibration, ROC/PR, predicted-vs-actual) — optional [explain] extra for SHAP; diagnostic plots need no extra deps (ADR 0020)
Deep learning	PyTorch (declarative nn.Module codegen seam; not a runtime dependency)
Visualization	Plotly
Custom code (emergentflow/script/)	User-defined def transform(value) Python nodes compiled and executed in-process
GenAI / LLM (emergentflow/llm/, emergentflow/eval/)	LiteLLM gateway behind an injected client seam ([llm] extra), replay/record fixtures for offline CI, Prompt Lab eval/label/export — ADR 0017. Multi-agent orchestration (LangGraph) is planned, not yet implemented (Epic 10).
Diagnostics	YData-Profiling

---

Roadmap

Development is structured into three iterative phases. The full epic-level decomposition lives in planning_docs/technical_roadmap.md.

Phase 1 — Foundation (SDK + static canvas) — done

The open-source Python SDK and graph IR, the codegen engine, the infinite-canvas frontend, a broad node library (data/clean/stats/viz/ml, full scikit-learn adapter), structural + nominal type validation, and save/load + export. Delivered: a canvas that maps a node graph to flawless, downloadable Python, plus a declarative nn.Module codegen seam for PyTorch architectures.

Phase 2 — Living Bridge (local reactive backend, bundled) — done

A local FastAPI/Uvicorn server bundled in the package (emergentflow serve) that runs ef.execute(ir) in-process, streams progress over SSE, DAG-aware incrementally caches results on disk, and renders results back into the canvas per node. The enterprise build-out — Celery/sandboxed/distributed execution, Redis + object-store caching, auth/multi-tenancy/deploy — stays deferred to the gated hosted product (ADR 0013 / §A6).

Phase 3 — Frontier (Deep Learning, GenAI & agents) — in progress

Delivered so far: LLM/Prompt Lab nodes with an injected, replayable client seam and an eval/label/export workflow (Epic 9); data-warehouse connectors as a second injected-client seam (Epic 13); agent collaboration with an in-app chat panel and multi-backend agent adapters (Claude, Gemini, Codex, OpenCode) — an AI agent proposes GraphMutations that a human reviews on the same canvas via ghost-diffs, gates, and review threads (Epic 14, ADR 0019); dedicated per-family statistical modeling nodes (linear regression, GLM, GAM, mixed models, Bayesian); feature transformation nodes (scaling, encoding, discretization, feature generation); SHAP-based model explainability with diagnostic plots (ADR 0020); custom-code nodes for user-defined Python transforms; markdown notes for graph annotation; a unified connection manager for warehouse, LLM, and agent profiles; and canvas UX improvements including CodeMirror 6 code editing, column-aware parameter widgets, and expandable inspector panels. Still ahead: visual PyTorch composition with real-time tensor-shape resolution, multi-step agentic orchestration / LangGraph codegen (Epic 10), RAG (Epic 11), and real-time multiplayer editing, with the IR designed CRDT-ready from the start.

---

Repository Layout

emergent-flow/
├── emergentflow/               # Python SDK source
│   ├── ir/                   # Graph intermediate representation (schema, serialization, migrations)
│   ├── nodes/                # Node contract, registry, catalog export, and reference examples
│   ├── codegen/              # Whole-graph compiler + reference executor (compile_to_code / execute)
│   ├── types/                # Type catalog, compatibility rules, rules-as-data artifact
│   ├── data/                 # ef.data (+ data/warehouse/: DuckDB/Postgres/BigQuery/Redshift adapters, query builder)
│   ├── clean, stats, ml, viz, reports/   # Reference node-family SDK wrappers
│   ├── explain/              # SHAP-based model explainability and diagnostic plots (ADR 0020, [explain] extra)
│   ├── script/               # Custom-code node: user-defined Python transform functions
│   ├── connections/          # Unified connection-profile store (warehouse + LLM + agent, TOML-backed)
│   ├── llm/                   # Injected LLM client seam (Gateway/Replay), prompt templating, budget/pricing
│   ├── eval/                  # Prompt Lab eval run / label / export
│   ├── collab/                # Agent collaboration: sessions, chat, agent adapters, GraphMutation, gates, review, MCP (Epic 14)
│   ├── server/                # Local FastAPI/Uvicorn server (ef.* in-process) — `emergentflow serve`
│   ├── clients.py             # Injected-client bundle threaded through execute()/compiled main()
│   ├── cli.py                 # `emergentflow` console entry point (serve / lab)
│   └── api.py                 # @public_op decorator + inspectable-return contract
├── ui/                        # TypeScript/React canvas (Vite), bundled into the wheel — never imports emergentflow
│   └── src/                   # canvas, palette, inspector, exec, session (chat/collab UX), promptlab, connections, store
├── agents/                    # Persona files for AI agents collaborating on a graph over HTTP (Epic 14)
├── docs/
│   ├── adr/                  # Architecture Decision Records (foundational decisions)
│   ├── runbook.md            # How to install, launch, and use the app (UI, API, CLI)
│   ├── acceptance-demo.md    # End-to-end worked pipeline
│   ├── agent-integration.md, bring-your-own-key.md  # Agent collaboration protocol + LLM credentials
│   ├── node-contract-spec.md, node-registry.md, node-catalog-artifact.md, authoring-a-node.md
│   ├── codegen-compiler.md, codegen-executor.md, codegen-traversal.md, codegen-declarative.md, how-codegen-works.md
│   ├── ir-spec.md, ir-serialization-format.md, ir-migrations.md, result-payload-contract.md
│   ├── type-system-spec.md, connection-validation.md
│   ├── package-layout.md, public-api-conventions.md, sdk-design-philosophy.md, ui-server-boundary.md
│   ├── versioning-and-releases.md, licensing-and-dependencies.md, open-core-boundary.md
│   └── stats-viz-design.md
├── epics/                     # Epic 1–14 story-level decomposition (README.md maps repo ↔ roadmap numbering)
├── examples/                  # Ready-made IR graphs + acceptance demos (sklearn, stats/viz, data connectors, agent collab)
├── planning_docs/
│   ├── proposal.md            # Product vision & market mapping
│   └── technical_roadmap.md   # Engineering decomposition (epics, phases, decisions)
├── schema/                    # Exported JSON Schemas (diagnostics, connection-validation rules)
├── scripts/                   # export_ui_contracts.py, check_ui_boundary.py (CI-enforced UI/SDK boundary)
├── .github/workflows/         # CI (Python lint/type/test, UI build/test, warehouse driver integration) and release pipelines
├── pyproject.toml             # Packaging, dependencies, tool config (ruff, mypy, pytest)
├── uv.lock                    # Pinned, reproducible dependency lockfile
└── README.md

---

Documentation

• Runbook — install, launch (emergentflow serve), use the canvas, drive the REST/SSE API directly, develop ui/, and run the full local CI gate set.
• The Acceptance Demo — the end-to-end pipeline that shows what the app can do today.
• Agent Integration and the agents/emergent-flow-collaborator.md persona — how an AI agent (Claude Code, Shards, or any HTTP client) joins a session and proposes GraphMutations alongside a human, per ADR 0019.
• Bring Your Own Key — how LLM/warehouse credentials are supplied (env-var names only; never embedded in the IR — ADR 0017/0018).
• Package Layout & Namespace Conventions — the emergentflow / ef package structure and namespaces.
• Public API Conventions — naming, signatures, and the serializable + inspectable return-object contract every wrapper must meet.
• SDK Design Philosophy — the thin / deterministic / pure rules and the @ef.public_op runtime check that enforces them.
• Codegen Engine — the whole-graph compile_to_code compiler and reference execute interpreter, plus the declarative nn.Module seam, traversal, and executor internals.
• How Codegen Works — an overview of the codegen pipeline, the two paradigms, and the golden / equivalence quality gates that back the "what you see runs" promise.
• The IR Spec, Serialization Format, and Migrations — the graph intermediate representation and its wire-format evolution policy.
• Result Payload Contract — the schema every node's execute output and the server's /execute response conform to.
• The Type System & Connection Validation — the nominal type model, compatibility + cardinality rules, whole-graph inference, ef.validate/Diagnostics, the strictness policy, and the shared codegen/execute gate.
• Connection Validation: Rules as a Portable Artifact — the type rules + Diagnostics schema shipped to the frontend canvas as data, and the frontend-vs-SDK authority model.
• UI ↔ Server Boundary — the contract artifacts (ir.schema.json, the node catalog, mutation/session-event schemas) that are the only thing crossing the ui/ ↔ emergentflow/ wall, and the CI check that enforces it.
• Node Catalog Artifact — how the palette/config-panel catalog is exported as data for the canvas.
• Versioning & Releases — Semantic Versioning policy and the tag-driven release process.
• Architecture Decision Records — the foundational §A decisions.
• How to Author a Node — the node-definition contract in practice.

The SDK uses uv for reproducible installs (uv sync), ruff for lint + format, and mypy for type-checking; the canvas uses npm, eslint, tsc, and vitest. CI runs the Python gates on 3.11/3.12, a separate ui/ job (lint, typecheck, build, test), and a driver-integration job against a live Postgres for the warehouse adapters.

---

License

The Emergent Flow SDK in this repository is licensed under the Apache License 2.0. Emergent Flow follows an open-core model: the Python SDK is open source, while the collaborative platform (visual canvas, real-time multiplayer, hosting, and enterprise features) is a separate proprietary product. See Open-Core Boundary for what is SDK vs. platform-only, and Dependency Licensing & Compatibility for the dependency-license policy.

Contributions are accepted under a license-grant CLA — see CONTRIBUTING.md.