Skip to content

README.md

Latest commit

 

History

History
258 lines (197 loc) · 17.6 KB

README.md

File metadata and controls

258 lines (197 loc) · 17.6 KB
Raw

GABRIEL

GABRIEL (Generalized Attribute Based Ratings Information Extraction Library) turns messy qualitative corpora into analysis-ready datasets with GPT. It handles prompting, batching, retries, checkpointing, and audit trails so you can treat “ask the model” workflows like any other measurement instrument. From rating rhetoric across a million speeches to matching product catalogs, you focus on the research question while GABRIEL handles the operations.

📓 Tutorial notebook (start here!): GABRIEL Colab notebook — also available as gabriel_tutorial_notebook.ipynb in this repo if you’d like to download and run it locally.

🖥️ Slides: Presentation deck

You can install the GABRIEL Python library from PyPI with pip install openai-gabriel and then import gabriel.

Compatibility note: max_output_tokens is deprecated in GABRIEL. Existing calls remain valid, but the value is ignored and emits a FutureWarning; GABRIEL leaves the output budget to the selected model.

Using a coding agent? Read AGENTS.md. It is the repository's shared guide to model choice, data shape, scale, checkpoints, and quality assurance.

Read our blog post here, our paper with validation experiments and applied examples here, and submit feedback / bugs / feature requests here.

Table of contents

Why GABRIEL?

Most of the evidence social scientists and analysts care about lives in unstructured formats: interviews, speeches, transcripts, product photos, archival scans. Modern GPT models can judge attributes, extract facts, and reason about this material with high fidelity, but building robust pipelines is still tedious. GABRIEL provides:

  • 🧠 Human-level comprehension on demand – express the attribute the way you would brief a human coder; GABRIEL packages the prompt, context, and retries for you.
  • 📊 Quantitative outputs – ratings (0–100), pairwise comparisons, classifications, and structured extractions return as tidy DataFrames with reproducible settings.
  • ⚙️ Operational tooling – automatic parallelism (hundreds of concurrent calls), resumable runs, raw response logs, and helper UIs make it safe to scale to very large corpora.
  • 🧰 Extensibility – swap instructions with additional_instructions, bring your own templates, or drop down to gabriel.whatever + custom response_fn for bespoke prompts while still reusing the infrastructure.

The tutorial notebook walks through these ideas step-by-step—from setting up an API key to running multimodal analyses—so skim this README, then dive into the notebook for the full guided tour.

What can you do with GABRIEL?

A) Measure attributes on qualitative data

Function Purpose & Output Scale Example Use
gabriel.rate Asks GPT to score each text / image / audio / item on natural language attributes. Output = 0-100 rating. Measure “populist rhetoric” in a speech; “toxicity” of tweets; “luxury” in ad images.
gabriel.rank Pairwise comparisons yield Bradley–Terry attribute ratings. Output = model-derived z-scores standardized within connected comparison components. Rank technologies by “bulkiness” or artworks by “fine brushwork”.
gabriel.classify Classifies texts / images / audio / items on whether provided labels apply. Output = one or more classes per item. Tag news articles, product photos, or interview clips into topical categories.
gabriel.extract Structured fact or entity extraction. One source can return one row or expand into several extracted-entity rows. Extract every product and its name, price, description, and color from each catalog page.
gabriel.discover Discovers natural language features which discriminate two classes of data. Identify what distinguishes 5 star vs. 1 star reviews or successful vs. failed startups.

B) Clean data

Function Purpose & Output Scale Example Use
gabriel.merge Creates crosswalks. Output = merged table with GPT-matched identifiers. Match two distinct job title directories; link patent titles to product names.
gabriel.deduplicate Detects conceptual duplicates. Maps all duplicates to one representative term. Collapse “F-18”, “Super Hornet Fighter Jet”, “f-18 hornet” into “F-18”.
gabriel.filter High-throughput boolean screening. Outputs items which meet natural language condition. Subset 18M Wikipedia titles to only technologies.
gabriel.deidentify Replaces PII with realistic, consistent fake PII. Outputs anonymized text + mapping. Replace names, employers, addresses before sharing interview corpora.

C) Helper tools

Function Purpose & Output Scale Example Use
gabriel.codify Passage coding: highlights snippets in text that match qualitative codes. Flag sentences about “economic insecurity” in speeches; “stressors” mentioned in interview.
gabriel.compare Identifies similarities / differences between paired items. Output = list of differences. Contrast op-eds from different districts; compare two ad campaigns.
gabriel.bucket Proposes a compact taxonomy from many terms. Output = bucket names and definitions, not row-level assignments. Develop candidate categories for technologies, artworks, or HR complaints, then review and apply them downstream.
gabriel.seed Enforces a representative distribution / diversity of seeds. Initialize unique personas that match US population distribution.
gabriel.poll Seeds personas, expands them into full biographies, and surveys them. Simulate a synthetic opinion poll on policy, values, and open-ended attitudes.
gabriel.ideate Generates and filters candidate scientific theories for further evaluation. Procure novel theories on inflation for potential research.
gabriel.debias Post-process measurements to remove inference bias. Ensure GPT isn't guessing climate opinions in speeches based on general political lean.
gabriel.load Prepares a folder of text / image / audio files into a spreadsheet for use in GABRIEL. Image directory converted into spreadsheet of file paths.
gabriel.view UI to view sample texts with ratings / passage coding. Spot-check classify / rating outputs; view coded passages.
gabriel.paraphrase Rewrites texts consistently per instructions. Summarize earnings call transcripts to remove company specifics.
gabriel.whatever Run any GPT prompts, but leverage GABRIEL's parallelization / checkpointing. Any set of prompts; slots into any pipeline.

Installation

pip install openai-gabriel

# or install directly from GitHub
pip install \
  --force-reinstall \
  git+https://github.com/openai/GABRIEL.git@main

# then run import gabriel

Before running a GABRIEL call, declare your GPT API key:

export OPENAI_API_KEY="sk-..."

In Colab, store the key in Secrets under OPENAI_API_KEY and read it with google.colab.userdata; do not paste a real key into notebook source, chat, screenshots, or Git. The tutorial shows the complete setup.

Using GABRIEL with coding agents

The repository's AGENTS.md is a concise operating guide for coding agents. It explains task selection, model routing, checkpointing, multimodal inputs, and QA. Agents should also read gabriel_tutorial_notebook.ipynb closely before designing a workflow; it is the broadest guide to what the library can do. Three rules are especially important:

  • Verify model slugs live. At this documentation's last review, the exact general models were gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna, and audio used gpt-audio-1.5. Model IDs change; check the official model catalog and pricing, plus the models actually provisioned for your organization or gateway, instead of guessing or copying an old example. Luna is often a sensible starting point for many applications and large-scale work. For unusually subtle, knowledge-heavy, multi-constraint, or writing-sensitive work, pilot Terra and Sol on representative examples and use the smallest model that reliably meets the research standard.
  • Generally keep n_parallels=650, including for small tests. It is a ceiling, not a fixed worker count. GABRIEL ramps up, reads rate-limit signals, retries transient failures, and lowers actual concurrency when needed. Temporary retries and a few slow stragglers are normal: while the run is progressing, let it finish. Raising the ceiling can help a large workload when provisioned limits support more throughput; lowering it may help a known deployment constraint or persistent failure. A small test does not require a lower ceiling.
  • Pilot, checkpoint, then scale. Test a representative sample, inspect raw responses and output shape, and keep reset_files=False to resume a compatible run. Use a new save_dir when the model, prompt, fields, or labels change.

Quick start

The tutorial notebook walks through many complete projects; here’s the minimal rating flow the notebook starts with. Paste this into Colab or a notebook cell so you can use await directly:

import os
import pandas as pd

import gabriel

PATH = os.path.expanduser("~/Documents/gabriel_runs")
toy_data = pd.DataFrame(
    {
        "entity": [
            "turkey",
            "pumpkin pie",
            "green bean casserole",
            "cornbread",
        ]
    }
)

attributes = {
    "savory taste": "How savory the dish is",
    "sweet taste": "Dessert-like sweetness",
    "tangy taste": "Notes of tartness or acidity",
}

rate_results = await gabriel.rate(
    toy_data,
    column_name="entity",
    attributes=attributes,
    save_dir=os.path.join(PATH, "toy_rate"),
    model="gpt-5.6-luna",
    n_parallels=650,  # normally leave this default unchanged
    n_runs=1,
    modality="entity",
    reset_files=True,
)
rate_results.head()

The helper returns a pandas.DataFrame with one column per attribute and writes raw model responses + configs to save_dir. Running the same code in a plain Python script just requires wrapping the coroutine with asyncio.run(...).

Task highlights

The tutorial notebook covers full projects end-to-end. The list below matches its main use cases so you can jump to the right helper quickly.

Measurement primitives

  • gabriel.rate – assign 0–100 scores per attribute across text, entities, images, audio, or web-sourced context.
  • gabriel.rank – pairwise tournaments fitted with a Bradley–Terry model, with symmetric pseudo-comparison regularization on observed edges when learning_rate > 0; a zero value requires Ford's directed strong-connectivity condition within each nontrivial observed component for a finite fit. Draws are encoded as half-wins in a binary Bradley–Terry working objective (one comparison total), not as a three-outcome tie likelihood. insufficient_signal_policy="tie" makes the modeling assumption that missing direct evidence is equality; use "abstain" to condition fitting on retained judgments, noting that informative abstention can still bias estimates. In non-recursive runs, z-scores are standardized separately within connected comparison components (with a neutral zero fallback when a component has no score dispersion), and _component columns identify which values are mutually comparable; isolated items remain unranked with NaN scores and standard errors. <attribute>_se is an asymptotic, model-based sandwich standard error for the component-centered raw log-skill <attribute>_raw, computed while treating the realized comparison graph as fixed. Observed comparison weights determine the meat under the binary-BT working variance, while fixed pseudo-comparisons add curvature to the bread. The standard error excludes shrinkage bias, shared-judge dependence, misspecification, adaptive-selection uncertainty, and uncertainty for the reported z-score. A non-recursive <file_name>_diagnostics.csv sidecar records decoded outcome counts, effective comparison weight, graph components, isolates, and the count of finite standard errors. Resume metadata fingerprints the effective judge settings and input payloads; judge_version is required for custom judges and must change whenever their hidden logic changes. Planned rounds and Batch jobs are recovered transactionally rather than silently mixed or resubmitted.
  • gabriel.classify – single- or multi-label tagging with label definitions and consensus columns.
  • gabriel.extract – extract typed fields or multiple entities from each passage, page, image, PDF, or audio item; one source may intentionally produce several rows.
  • gabriel.discover – contrast two labeled corpora to learn discriminating features.

Qualitative coding and review

  • gabriel.codify highlights snippets that match qualitative codes and pairs with gabriel.view for UI-based auditing.
  • gabriel.compare contrasts paired items (drafts, policies, campaigns) with concise differences/similarities.
  • gabriel.bucket proposes bucket names and definitions from a large term universe; review that taxonomy, then use classify or another mapping step when row-level assignments are needed.

Data prep and cleanup

  • gabriel.load converts folders of media into spreadsheets with clean IDs and file paths.
  • gabriel.merge / gabriel.deduplicate produce fuzzy joins and de-duplicated lists using embeddings plus GPT checks.
  • gabriel.filter screens large candidate lists with natural-language conditions.
  • gabriel.deidentify replaces PII with realistic stand-ins to protect privacy.

Ideation and custom prompts

  • gabriel.ideate and gabriel.seed generate diverse candidates before deeper measurement.
  • gabriel.whatever runs bespoke prompts (with optional web search or custom response_fn) while reusing retries, logging, and checkpointing.

Multimodal data and web search

Set modality to text, entity, pdf, image, audio, or web on any measurement helper. Pair gabriel.load with folders of media to build the right DataFrame, and use web_search=True when GPT should gather context before rating or extracting. The tutorial’s county-level example shows how to chain web search → rating → mapping in one flow.

For image understanding, pass image_detail="low", "high", "original", or "auto"; this is distinct from image-generation quality. Audio inputs need an audio-capable model and are routed through Chat Completions rather than the Responses API. At this documentation's 2026-07-13 review, use gpt-audio-1.5; gpt-audio-mini is deprecated. Verify the current audio slug before running an old example. Treat GABRIEL's pre-run cost estimate as rough for audio, tool calls, and very long contexts; consult live API pricing for the applicable modality and processing tier.

Custom prompts and model routing

  • Add clarifications with additional_instructions (e.g., mandate mutually exclusive labels).
  • Swap in your own Jinja template_path while keeping retries and checkpoints.
  • Drop to gabriel.whatever for fully custom prompts, attachments, or routing logic.
  • Treat every model name in an example as a dated snapshot; verify the exact current slug and endpoint support before overriding a default.

Saving, logging, and resuming

Each run expands save_dir (tilde and environment variables supported), writes cleaned outputs, and checkpoints raw calls in task-specific files such as *_raw_responses.csv. Leave reset_files=False to resume a compatible partial run. When the research specification changes, use a fresh directory or pass reset_files=True intentionally so old responses are not mixed with a new prompt. Keep large generated outputs and private datasets outside the source repository. gabriel.view reads results for quick spot checks, and helpers like gabriel.utils.mapmaker.MapMaker can consume the same files downstream.

Development and testing

Install development extras and run tests:

pip install -e .[dev]
pytest

Tests rely on the built-in dummy responses, so no API key is necessary. Linting and type checks (ruff, mypy) are included in the dev extras.

Citation

If you use GABRIEL in your research, please cite:

  • Asirvatham, H., Mokski, E., and Shleifer, A. (2026). GPT as a Measurement Tool. NBER Working Paper No. 34834.

  • Asirvatham, H. and Mokski, E. (2026). GABRIEL: Generalized Attribute-Based Ratings Information Extraction Library (software). GitHub repository: https://github.com/openai/GABRIEL

License

GABRIEL is released under the Apache 2.0 License. See LICENSE.