va-dat — Vision Aid Digital Accessibility Testing

LLM-powered tools that analyze websites for WCAG accessibility issues and generate structured remediation reports. A Computing for Good course project at Georgia Tech (OMSCS) partnered with the Vision Aid Digital Accessibility Testing Team.

There are two ways to use it: a web app (paste a URL or HTML, get findings and a CSV) and a CLI pipeline for batch or scripted runs. Both share the same analysis code.

How It Works

The pipeline takes a raw HTML file and produces targeted accessibility findings through four steps:

HTML file (e.g. 1.9 MB)
    │
    │  Step 0 — Programmatic Checks (no API cost)
    │  Rule-based detection of missing alt, empty links, duplicate IDs, etc.
    │
    │  Step 1 — Extract
    │  Three extractors parse the HTML into structured JSON payloads,
    │  discarding layout noise (CSS, scripts, divs) and keeping only
    │  semantically relevant content. ~92% token reduction.
    │
    ├── semantic_checklist_01.py  → headings, links, landmarks, tables, iframes
    ├── forms_checklist_02.py    → form fields, labels, groups, instructions
    └── nontext_checklist_03.py  → images, SVGs, icon fonts, media
                                    Combined: ~39k tokens (from ~487k)
    │
    │  Step 2 — Slice & Call
    │  Each payload is sliced into targeted pieces. Each slice is paired
    │  with a focused prompt template and sent to the LLM individually.
    │  Up to 18 element-specific calls + 3 optional summary calls.
    │
    │  Step 3 — Save
    │  Raw JSON results are saved per-prompt for downstream processing.
    │
    │  Step 4 — Report
    │  The report generator reads all saved results, normalizes findings
    │  from both programmatic and LLM sources, and writes a unified CSV.

Repository Structure

├── index.html                                   # Web UI + team site (all markup/CSS/JS inline)
├── styles.css
│
├── entry_points/                                # Entry points
│   ├── api_server.py                            # Serves the site and the /api/* audit endpoints
│   ├── run_pipeline.py                          # Runs the full pipeline from the CLI
│   └── generate_report.py                       # Combines findings into unified CSV report
│
├── processing_scripts/
│   ├── llm/                                     # Modular prompt system + templates
│   │   ├── registry.py                          #   Maps each evaluation task to its template + slicer
│   │   ├── templates.py                         #   Parses .txt prompt files, fills {payload} placeholders
│   │   ├── slicers.py                           #   Extracts targeted JSON slices from extractor payloads
│   │   ├── semantic_checklist_01.txt            #   7 prompts for semantic structure
│   │   ├── forms_checklist_02.txt               #   6 prompts for form accessibility
│   │   └── nontext_checklist_03.txt             #   8 prompts for non-text content
│   │
│   ├── llm_client/                              # Standalone Claude API client (Andrew)
│   │   ├── client.py                            #   API wrapper
│   │   ├── prompt_loader.py                     #   Prompt loading utilities
│   │   └── runner.py                            #   End-to-end audit runner
│   │
│   ├── llm_preprocessing/                       # HTML → structured JSON extractors
│   │   ├── semantic_checklist_01.py             #   Headings, links, landmarks, tables, iframes
│   │   ├── forms_checklist_02.py                #   Form fields, label associations, groups
│   │   └── nontext_checklist_03.py              #   Images, SVGs, icon fonts, media
│   │
│   ├── programmatic/                            # Rule-based checks (no LLM needed)
│   │   ├── semantic_checklist_01.py             #   Semantic structure checks
│   │   ├── forms_checklist_02.py                #   Form accessibility checks
│   │   └── nontext_checklist_03.py              #   Non-text content checks
│   │
│   └── docs/pipeline.md                         # Pipeline architecture documentation
│
├── vision_aid/ingestion/
│   ├── file_crawler.py                         # fetch_page / fetch_pages_nested (used by the server)
│   └── pull_html.py                            # Standalone HTML download helper
│
├── Dockerfile                                  # Multi-stage uv build → runtime image
├── docker-compose.yml                          # Local run (Coolify deploys the image directly)
├── DEPLOY.md                                   # Coolify deployment notes
│
├── .github/workflows/
│   ├── ci.yml                                  # On PR to main: deps, imports, pipeline, image smoke test
│   └── publish.yml                             # On push to main: build → GHCR → trigger Coolify
│
├── test_files/                                 # HTML files to analyze
│   ├── home.html                               #   visionaid.org homepage (~1.9 MB)
│   └── dat_visionaid_home.html                 #   Smaller trimmed variant (~143 KB)
│
├── semantic_checklist/                         # Source of truth: Deque WCAG checklist PDFs
│   ├── 01-semantic-checklist.pdf
│   ├── 02-forms-checklist.pdf
│   └── 03-nontext-checklist.pdf
│
├── pipeline_walkthrough.ipynb                  # Colab-compatible step-by-step pipeline notebook
│
├── docs/                                       # Architecture documentation
│   └── modular-prompts-plan.md                 #   Full architectural plan
│
├── reports/                                    # Raw JSON output from standalone llm_client runs
│
├── test_results/
│   ├── chatgpt/                                # Legacy ChatGPT testing results
│   └── claude/                                 # Pipeline-generated CSV reports
│       └── report_YYYY-MM-DD.csv
│
└── output/                                     # Generated at runtime (not committed)
    ├── manifest.json
    ├── programmatic_findings.json
    ├── payloads/
    └── prompts/

Architecture Deep-Dive

Extractors (existing code)

Each extractor has an extract(file_path) function that parses HTML with BeautifulSoup and returns a structured dict:

Extractor	Focus	Output tokens (visionaid.org)
semantic_checklist_01.py	Page title, headings, links, landmarks, tables, iframes	~17,600
forms_checklist_02.py	Form fields with label source, instructions, required flags	~2,500
nontext_checklist_03.py	Images (4 categories), SVGs, icon fonts, video/audio	~19,200

These files live in processing_scripts/llm_preprocessing/ and were authored by ahildebrandt3 and Andrew Yin. They should not need modification unless a new checklist (CL04+) is added.

Prompt Registry Pattern (new)

The core of the modular system is in processing_scripts/llm/:

• registry.py — Defines 21 PromptSpec dataclass entries, each linking a prompt name to its template file, slicer function, WCAG criteria, and output type. This is the single source of truth for what the pipeline evaluates.

• slicers.py — Contains one function per prompt (e.g., slice_headings(), slice_flagged_links()) that extracts exactly the data that prompt needs from the full extractor payload. This is what achieves the token reduction.

• templates.py — Parses the .txt prompt template files (which contain multiple numbered prompts separated by dashed headers) and fills in the {payload} placeholder with the sliced JSON at runtime.

Pipeline Orchestrator (new)

entry_points/run_pipeline.py ties everything together:

1. Runs the three extractors to get structured payloads
2. Runs programmatic checks on the CL01 payload
3. Iterates over the prompt registry, slicing payloads and assembling prompts
4. Calls the Anthropic API for each non-empty prompt (or saves dry-run output)
5. Writes a manifest with token counts, timing, and cost data

Report Generator (new)

entry_points/generate_report.py reads the pipeline output and produces a flat CSV:

1. Loads manifest.json for run metadata (date, model)
2. Normalizes programmatic_findings.json (59 rule-based issues) into report rows
3. For each output/prompts/*.json, applies a prompt-specific normalizer that understands the response schema and extracts issues
4. Assigns sequential IDs and writes to test_results/claude/report_YYYY-MM-DD.csv

The normalizer registry mirrors the prompt registry — one normalizer function per prompt type that knows how to detect issues in that prompt's response shape.

How to Extend

Adding a new prompt type

1. Slicer — Add a function in processing_scripts/llm/slicers.py that extracts the relevant data from the extractor payload
2. Template — Add a new numbered prompt section to the appropriate .txt file in processing_scripts/llm/
3. PromptSpec — Add an entry in processing_scripts/llm/registry.py linking the slicer, template, and WCAG criteria
4. Normalizer — Add a normalizer function in entry_points/generate_report.py and register it in the NORMALIZERS dict

Adding a new extractor/checklist (CL04+)

1. Create a new extractor in processing_scripts/llm_preprocessing/ with an extract(file_path) function
2. Create corresponding prompt templates in processing_scripts/llm/
3. Add slicer functions in processing_scripts/llm/slicers.py
4. Register new PromptSpec entries in processing_scripts/llm/registry.py
5. Add normalizers in entry_points/generate_report.py
6. Update entry_points/run_pipeline.py to call the new extractor

Setup

This project uses uv. It installs the right Python version itself, so there is no separate Python install or venv step.

# Install uv once (see the uv docs for Windows/other options)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Create the environment and install exactly the locked dependencies
uv sync

Then either prefix commands with uv run, or activate the environment:

uv run python entry_points/run_pipeline.py --help
# or
source .venv/bin/activate        # Linux/macOS
# or: .venv\Scripts\activate      # Windows

uv.lock pins every dependency, including transitive ones, so everyone gets an identical environment. To change a dependency, edit pyproject.toml and run uv lock, then regenerate the pip fallback below.

Using pip instead (graders, Colab, or no uv available)

requirements.txt is a generated export of uv.lock — do not edit it by hand; regenerate it with the command in its header. Python 3.11+ required.

python -m venv venv
source venv/bin/activate        # Linux/macOS
# or: venv\Scripts\activate     # Windows

pip install -r requirements.txt
pip install -e . --no-deps

If python -m venv fails with an ensurepip error, your Python is missing its venv module (sudo apt install python3-venv on Debian/Ubuntu). uv avoids this problem entirely.

Create a .env file with your provider API key(s) (only needed for live runs, not dry-run):

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
GEMINI_API_KEY=AIza...

.env is gitignored. Never commit a key.

A run with no resolvable key silently becomes a dry run — programmatic checks only, no LLM findings, no CSV, and still a 200 OK from the web app. The only signal is summary.dry_run in the response.

Running the Web App

uv run python entry_points/api_server.py      # http://localhost:8000

Serves the UI and the audit API from one process. Users can paste their own API key into the form instead of configuring one server-side; per-request keys take priority over the environment.

The audit endpoints stream NDJSON — progress events, one JSON object per line, then a final {"type":"result"} object. Parsing that body with a single res.json() fails; the front end branches on content type.

To run it in a container:

docker compose up --build                     # http://localhost:8000
HOST_PORT=8789 docker compose up --build      # if 8000 is taken

Deployment

Merges to main build the image, push it to ghcr.io/c4g/va-dat, and trigger a Coolify deploy to https://va-dat.c4g.dev. See DEPLOY.md — in particular the proxy settings, since response buffering breaks the progress stream and short read timeouts cut off long audits.

Note that Coolify runs the image; the hardening in docker-compose.yml (read_only, tmpfs) applies to local runs only.

Continuous Integration

.github/workflows/ci.yml runs on every PR to main and needs no API key — everything it does is free:

• uv.lock is in sync with pyproject.toml, and requirements.txt matches the lock
• entry points import
• a full pipeline dry run, asserting prompts generated, findings found, and zero tokens consumed
• index.html's inline JavaScript parses
• the Docker image builds, becomes healthy, serves the site, and returns a valid NDJSON audit

Running the Pipeline

Dry run (no API calls, no cost)

Generates all prompts and saves them as JSON files so you can inspect them before spending money:

uv run python entry_points/run_pipeline.py --html test_files/dat_visionaid_home.html --dry-run

Live run

Sends prompts to the LLM and saves responses:

uv run python entry_points/run_pipeline.py --html test_files/home.html

Generate report

After a live run, combine all findings into a single CSV:

uv run python entry_points/generate_report.py
uv run python entry_points/generate_report.py --output-dir ./output --report-dir ./test_results/claude/

Pipeline options

Flag	Default	Description
--html	(required)	Path to the HTML file to analyze
--output-dir	./output	Directory for results
--model	claude-sonnet-5	Anthropic model to use
--dry-run	off	Generate prompts without calling the API
--include-summaries	off	Include the 3 cross-cutting summary prompts
--show-cost	off	Print estimated dollar cost of the run based on model pricing
--env-file	.env	Path to environment file

Report generator options

Flag	Default	Description
--output-dir	./output	Directory containing pipeline output
--report-dir	./test_results/claude/	Directory to write the CSV report

Output Structure

Pipeline output (output/)

output/
├── manifest.json                 # Run metadata, token counts, prompt status
├── programmatic_findings.json    # Rule-based checker results (free)
├── payloads/                     # Raw extractor output (for inspection)
│   ├── cl01_payload.json
│   ├── cl02_payload.json
│   └── cl03_payload.json
└── prompts/                      # One file per prompt
    ├── page_title.json           # Contains prompt text, payload slice, and API response
    ├── heading_structure.json
    ├── link_clarity.json
    └── ...

Report output (test_results/claude/)

The report CSV has 13 columns matching the Vision Aid team's standard format:

Column	Description
ID	Sequential row number
element_name	HTML element (e.g. <img class="...">, <a> "link text")
browser_combination	Always N/A (static HTML analysis)
page_title	Page title from the analyzed HTML
issue_title	Short issue description
steps_to_reproduce	Element snippet or inspection steps
actual_result	What was found
expected_result	What WCAG requires
recommendation	Suggested fix
wcag_sc	WCAG success criterion (e.g. 1.1.1)
category	Issue category (e.g. Programmatic / Non-text Content)
log_date	Date of the pipeline run
reported_by	Programmatic or the LLM model string

Cost Estimate

For visionaid.org homepage (using Claude Sonnet):

Approach	Input tokens	Cost
Monolithic (entire HTML)	~487,000	~$1.52
Element-specific pipeline	~18,000	~$0.32

The pipeline skips prompts with empty payloads (e.g., no forms on the page = no form prompts), so actual cost varies by page content.

Attribution

Contributor	What they own	Key files
ahildebrandt3	Extractors, programmatic checkers (CL01–CL03), CL01 prompts	processing_scripts/llm_preprocessing/, processing_scripts/programmatic/
Andrew Yin	CL02 + CL03 extractors, CL02 + CL03 prompts, LLM client, pipeline docs	processing_scripts/llm_preprocessing/, processing_scripts/llm_client/, processing_scripts/llm/*.txt
nfulton99	HTML ingestion, packaging	vision_aid/ingestion/pull_html.py, pyproject.toml
ColeANiblett	Pipeline orchestration, prompt system, report generator	processing_scripts/llm/{registry,slicers,templates}.py, entry_points/, docs/