Turn any content into structured, actionable intelligence — in seconds.
PACE ingests content from 5 sources, runs 10 AI analyses in sequential batches, and delivers a structured report — all in a single click.
┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ YouTube │ │ PDF │ │ Article │ │ Audio │
│ Video │ │ Document │ │ (URL) │ │ File │
└──────┬──────┘ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘
│ │ │ │
└─────────┬─────────┴─────────┬─────────┘ │
│ │ │
▼ ▼ │
┌─────────────────────────────────────┐ │
│ Ingest → Clean → Chunk │◄───────────────────┘
└─────────────────┬───────────────────┘
│
▼
┌─────────────────────────────────────┐
│ Sequential Batch LLM Analysis │
│ A ──2s──► B ──2s──► C ──1s──► │
│ Synthesis │
│ (auto-fallback on failure) │
└─────────────────┬───────────────────┘
│
▼
┌─────────────────────┐
│ Report + Markdown │
│ PDF + EPUB + MD │
│ + Obsidian Export │
└─────────────────────┘
|
|
|
3 batches run sequentially with delays to avoid rate limits. Auto-fallback to individual steps on batch failure. |
LRU cache (50 entries, 1h TTL) — re-analyze the same content instantly. |
Built-in free tier via OpenCode Zen, or use your own key from 8 supported providers. |
|
Sidebar shows live connection status with a Test Connection button. Auto-detects provider from API key prefix. |
Markdown, PDF, EPUB, and Obsidian (YAML frontmatter + vault paths). |
Input sanitization, SSRF blocking, error sanitization, file validation. |
git clone https://github.com/AshayK003/PACE.git
cd PACE
python -m venv .venv
source .venv/bin/activate # macOS/Linux
# .venv\Scripts\activate # Windows
pip install -r requirements.txtstreamlit run app/main.pyOpens at http://localhost:8501
- Select a source tab (YouTube, PDF, Article, Audio, or Text)
- Paste a URL or drop a file
- Click Analyze
- Download as Markdown or PDF
PACE works out of the box with a built-in free tier. To use your own key:
- Open the sidebar → LLM Settings
- Select a provider or just paste your API key — the provider is auto-detected from the key prefix
- Select a model from the provider's list
- Click Test Connection to verify before analyzing
Your key is stored in session state only — never persisted to disk.
| Provider | Endpoint | Key Prefix | Models |
|---|---|---|---|
| OpenCode Zen (built-in free) | opencode.ai/zen/v1 |
— | deepseek-v4-flash-free, deepseek-v4-flash, deepseek-v3.1, gpt-4o-mini, gpt-4o, claude-sonnet-4, gemini-2.5-flash, llama-3.3-70b |
| Google Gemini (free) | generativelanguage.googleapis.com/v1beta/openai/ |
AIza |
gemini-2.5-flash, gemini-2.5-flash-lite |
| Groq (free) | api.groq.com/openai/v1 |
gsk_ |
llama-3.3-70b-versatile, llama-3.1-8b-instant, llama-4-scout-17b-16e-instruct, qwen3-32b, gpt-oss-120b |
| Cerebras (free) | api.cerebras.ai/v1 |
csk- |
llama-3.3-70b, gpt-oss-120b |
| OpenRouter (free) | openrouter.ai/api/v1 |
sk-or- |
deepseek/deepseek-r1-0528:free, deepseek/deepseek-chat-v3.1:free, qwen/qwen3-235b-a22b:free, meta-llama/llama-4-scout:free, meta-llama/llama-3.3-70b-instruct:free, google/gemma-4-31b-it:free, + more |
| Mistral (free) | api.mistral.ai/v1 |
— | mistral-small-latest, mistral-large-latest |
| DeepSeek | api.deepseek.com |
sk- |
deepseek-v4-flash, deepseek-chat |
| Custom | any OpenAI-compatible | — | user-specified |
| Variable | Required | Default | Description |
|---|---|---|---|
OPENCODE_ZEN_KEY |
No | Built-in free key | Default LLM API key |
Set via environment, .env, or .streamlit/secrets.toml.
app/
├── main.py # Streamlit entry, wires everything together
├── config.py # SourceType, safe_filename, build_export_path()
├── security.py # SSRF, rate limiting, input sanitization
├── ingestors/ # Content extraction (one per source)
│ ├── base.py # Abstract base class
│ ├── youtube.py # youtube-transcript-api + oEmbed title
│ ├── pdf.py # PyMuPDF4LLM + pdfplumber tables
│ ├── article.py # trafilatura + curl_cffi TLS impersonation + paywall bypass
│ └── audio.py # faster-whisper (local only)
├── processors/ # Text processing
│ ├── cleaner.py # Dedup, filler removal, URL stripping
│ └── chunker.py # semchunk semantic chunking
├── analyzers/ # LLM pipeline
│ ├── llm_client.py # OpenAI-compatible client + cache + rate limiter
│ ├── pipeline.py # Sequential batch execution (A→B→C with delays)
│ ├── prompts.py # All prompt templates (8 analysis sections)
│ ├── parser.py # ===SECTION=== delimited response parser
│ └── categorizer.py # LLM-based content categorization for export paths
├── output/ # Report generation
│ ├── markdown.py # Jinja2 → Markdown
│ ├── pdf.py # fpdf2 → PDF (lazy font loading)
│ ├── epub.py # ebooklib + mistletoe → EPUB
│ ├── obsidian.py # YAML frontmatter + vault paths
│ └── templates/ # Jinja2 templates
└── ui/ # Streamlit components
├── sidebar.py # LLM config, export path settings
└── components.py # Report display, download buttons
| Library | Why |
|---|---|
| Streamlit | Zero-cost deployment via Community Cloud, multi-user built-in |
| OpenAI SDK | Any OpenAI-compatible provider works |
| semchunk | Semantic chunking without embeddings |
| fpdf2 | Pure Python PDF, no system deps |
| PyMuPDF4LLM | 0.09s/page, 1GB RAM cloud-friendly |
| trafilatura | Highest F1-score article extractor in benchmarks |
| curl_cffi | TLS fingerprint impersonation bypasses Cloudflare/anti-bot |
| tenacity | Retry with exponential backoff |
# Run all 252 tests
pytest
# Verbose
pytest -v
# Specific module
pytest tests/test_analyzers.py -v
# Coverage
pytest --cov=app| Module | Tests | What's Covered |
|---|---|---|
test_analyzers.py |
30 | Pipeline, prompts, batching, cache |
test_security.py |
40 | SSRF, rate limiting, injection, file magic |
test_ingestors.py |
36 | YouTube, PDF, article (curl_cffi), audio |
test_output.py |
48 | Markdown, PDF, EPUB, Obsidian, Unicode, tables |
test_categorizer.py |
8 | LLM categorization for export paths |
test_cleaner.py |
20 | Filler removal, dedup, timestamps |
test_chunker.py |
10 | Semantic chunking, boundaries |
test_config.py |
24 | Filename utils, API key loading, export path |
test_parser.py |
9 | Batched response parsing |
test_integration.py |
16 | End-to-end pipeline resilience |
| Total | 252 |
- Push to GitHub
- Go to share.streamlit.io
- Select repo → branch
main→ entry pointapp/main.py - Add secrets:
OPENCODE_ZEN_KEY = "your-key-here"
# Local
streamlit run app/main.py
# Docker (contributions welcome)
docker build -t pace .
docker run -p 8501:8501 pace| Problem | Solution |
|---|---|
| Authentication error | Open sidebar → LLM Settings → select correct preset → enter your key |
| PDF garbled text | Scanned PDFs need OCR. Run locally with opendataloader-pdf |
| Audio tab disabled | faster-whisper requires local run. Not available on Streamlit Cloud |
| RequestBlocked (Gemini) | Google's safety filters blocked the content. Try a different model (gemini-1.5-flash) or provider |
| 429 Rate limit | Wait before retrying. Response cache avoids duplicate calls |
| Slow first request | PDF font downloads lazily. Cache warms after first analysis |
| SSRF blocks URL | Only public URLs allowed. Private IPs blocked for security |
We welcome contributions! Here's how:
- Fork the repo
- Create a feature branch:
git checkout -b feature/my-change - Make changes + add tests
- Run
pytest— all 252 tests must pass - Open a PR with a clear description
- New ingestor → Add
app/ingestors/my_source.py, inheritBaseIngestor, implementvalidate()+ingest() - New analysis step → Add prompt to
app/analyzers/prompts.py, register inALL_PROMPTS - New LLM preset → Edit
presetsdict inapp/ui/sidebar.py
PACE takes security seriously. See SECURITY.md for reporting vulnerabilities.
| Layer | Protection |
|---|---|
| Network | SSRF blocking (private IPs, cloud metadata, localhost) |
| File | Magic byte validation, 50MB size limit |
| API | Rate limiting (30 req/min), response caching |
| Input | 50K char truncation, prompt injection detection |
| Output | Error sanitization (no paths, no keys, no internals) |
MIT License — see LICENSE for details.
Built with care for the open-source community.
If you find PACE useful, consider giving it a star, or support the developer:
