accrue 1.5.2

Accrue
The enrichment pipeline engine.

Define a pipeline. Point it at your data. Get structured results. Accrue is a Python library for enriching datasets with LLMs. Compose multi-step pipelines, run them across hundreds to tens of thousands of rows, and get validated, structured output back -- with caching, retries, and parallel execution handled for you.

No platform. No markup. Just a pipeline you can version-control, iterate on, and reason about.

_{A 5,000-row run in Accrue Watch -- every step, model, field, and dollar, read live from the run log.}

from accrue import Pipeline, LLMStep

pipeline = Pipeline([
    LLMStep("analyze", fields={
        "market_size": "Estimate total addressable market in billions USD",
        "competition": {
            "prompt": "Rate competitive intensity with key competitors",
            "enum": ["Low", "Medium", "High"],
            "examples": ["High - Competes with AWS, Google Cloud"],
        },
        "growth_potential": {
            "prompt": "Assess 5-year growth trajectory",
            "type": "String",
            "format": "X% CAGR - reasoning",
        },
    })
])

result = pipeline.run(df)  # DataFrame in, DataFrame out
print(result.data.head())
print(f"Tokens used: {result.cost.total_tokens:,}")

Install

Requires Python 3.10+.

pip install accrue

Set your API key:

export OPENAI_API_KEY=sk-...

That's it. OpenAI is the default provider (zero config, structured outputs auto-enabled). Anthropic and Google are optional:

pip install accrue[anthropic]  # Claude
pip install accrue[google]     # Gemini

Claude Code Skill

If you use Claude Code, Accrue ships with a built-in /accrue skill that guides you through building pipelines interactively. It designs fields, picks models, estimates costs, and writes your script -- you just review and run.

> /accrue
> I have 500 companies in accounts.csv, I need to qualify them for ICP fit

The skill walks you through field design, model selection, pipeline architecture, and configuration before writing a production-ready script. See Using the Claude Code Skill for details.

Why Accrue

You have a spreadsheet of companies, leads, or entities. You need structured fields added to every row -- classifications, summaries, scores, extracted data. You could write a for loop and call the OpenAI API, but then you're building retry logic, rate limiting, caching, progress tracking, and crash recovery. You could use Clay, but you'd pay $500/month for something you can't version-control.

Accrue is the pipeline between a single API call and a full platform:

Quick Example

Chain steps together with depends_on. Use web_search() to ground LLM answers in live data:

from accrue import Pipeline, FunctionStep, LLMStep, web_search

pipeline = Pipeline([
    FunctionStep("research",
        fn=web_search("Research {company}: market position, competitors, recent news"),
        fields=["__web_context", "sources"],
    ),
    LLMStep("analyze",
        fields={
            "market_size": "Estimate TAM in billions USD",
            "competitors": {"prompt": "List top 3 competitors", "type": "List[String]"},
            "investment_thesis": "One-paragraph investment thesis",
        },
        depends_on=["research"],
    ),
])

result = pipeline.run(companies_df)

Watch It Run

Accrue Watch (accrue-ui) is a local dashboard for a running pipeline -- rows x steps as a live grid, every cell's values and errors, failures grouped by cause, cost by step. It reads the JSONL run log Accrue writes; there is nothing else to configure. The Overview tab is at the top of this page; this is the run grid:

_{Run grid, data view -- one field per step, filling in as rows complete; retries and failures show in place.}

Turn on the run log, then point the dashboard at it:

result = pipeline.run(companies_df, run_log=True)  # writes .accrue/runs/<run_id>.jsonl

pip install 'accrue[ui]'   # the dashboard, alongside the engine
accrue watch               # opens the latest run in your browser

accrue watch is a thin stub that delegates to accrue-ui. Add --pipeline module:attr to enable one-click retry of failed rows from the dashboard. Run log guide

Features

• Multi-step pipelines -- Chain LLM steps and function steps into a DAG with automatic dependency resolution and parallel execution. Quickstart

• Provider-agnostic -- OpenAI, Anthropic (with automatic prompt caching), and Google ship as adapters. Any OpenAI-compatible API works via base_url. Custom providers implement one async method. Providers guide

• 7-key field specs -- Control LLM output with prompt, type, format, enum, examples, bad_examples, and default. Drives structured outputs and Pydantic validation automatically. Field specs guide

• Caching and checkpointing -- SQLite input-hash cache auto-invalidates on prompt changes. Checkpointing saves after each step for crash recovery. Caching guide

• Batch API -- LLMStep(batch=True) for 50% cost savings via OpenAI and Anthropic batch endpoints. Cache-aware, auto-chunking, realtime fallback on failures. Batch guide

• Web search and grounding -- web_search() factory for search-then-analyze pipelines, or grounding=True for native provider web search with normalized citations. Web search guide

• Attachments -- LLMStep(attachments="documents") sends each row's PDFs to the model as document blocks instead of making you extract the text first. Cached by content, placeholdered in run logs. Anthropic today. Attachments guide

• Conditional steps -- run_if / skip_if predicates for per-row branching. Skipped rows get defaults, never hit the API. Conditional steps guide

• Hooks -- Typed lifecycle events for observability. Sync and async callables, never crash the pipeline. Hooks guide

• Run logs -- run(..., run_log=True) streams the run as append-only JSONL (schema v1): per-step and per-row events with status, errors, and token usage. Crash-safe, tail -f-able. Run log guide

• Run diffs -- accrue.compare(result_a, result_b) diffs two runs (e.g. before/after a prompt tweak) -- changed rows, per-field churn, distribution shift, and cost delta -- no labels needed. Compare guide

• Merging repeated runs -- accrue.merge_runs([run_a, run_b, run_c]) combines independent runs over the same rows into one answer per cell by majority vote, with per-cell agreement, the cells that need review, and cost summed across runs. Pluggable per-field resolvers for rules like "prefer the run with a verified quote". Merge guide

• provider_kwargs -- Escape hatch for provider-specific features (extended thinking, effort control, etc.) without waiting for first-class support.

Sweet Spot

Accrue is built for 100 to 50,000 rows -- too many for manual work or single-call tools, too few to justify big data infrastructure.

With batch=True, halve the API costs. Cached steps re-run in seconds.

Documentation

Contributing

git clone https://github.com/matt-house-e/accrue.git
cd accrue
pip install -e ".[dev]"
pytest

License

MIT