Sonario is a Windows desktop app with two jobs:
- Analyze Collection reads every supported file in one or more Windows or Google Drive folders and writes a report about what recurs across the collection.
- Summarizer turns a document, EPUB, YouTube link, or web page into a clear summary with downloadable Markdown and PDF output.
The app runs locally at http://127.0.0.1:5005. The default provider,
Qwen3.5 9B through Ollama, runs on your own machine with no account or API key.
The optional cloud provider is Groq – Qwen 3.6 27B, pinned to
qwen/qwen3.6-27b.
Timestamped transcript on the left, summary and follow-up questions on the right.
Installation, launcher reconstruction, release packaging, and Google Drive setup are documented in BUILD.md.
This repository reflects the currently tested Windows release:
- Groq uses Qwen 3.6 27B; the retired Scout-era Groq preset is not used anywhere.
- Groq structured calls use JSON Object Mode with larger retry budgets.
- If a model still returns malformed or truncated JSON, Sonario repairs or recovers the result instead of dropping the document. The progress log marks this as
JSON recovered. - Sonario opens in a dedicated maximized Edge or Chrome app window with browser sign-in and synchronization disabled for that private profile.
- Closing the Sonario window with X stops the exact Flask backend that opened it.
stop.batremains the manual fallback. - Google Drive access is optional, read-only, and starts only after the user deliberately connects or analyzes a Drive folder.
The X-close path was verified with a Windows shutdown diagnostic: the watcher exited,
port 5005 stopped listening, and no Sonario app.py process remained. Old
TIME_WAIT rows in netstat are normal closed connections, not a running server.
- Extract the complete Sonario release ZIP into an empty folder.
- Double-click
setup_all.batonce. - Launch Sonario from the desktop shortcut created by setup.
- Close the dedicated Sonario window with X when finished.
The release ZIP contains the local BAT launchers. BAT files are intentionally not stored in GitHub; their authoritative contents are kept in BUILD.md.
A clone does not contain BAT files. Recreate the required launchers from BUILD.md,
save them with CRLF line endings, then run setup_all.bat.
Demo collection:
Sample Documents/contains 25 fictional nested text files for testing recursive collection analysis. It contains no personal data.
Select a Windows folder, Google Drive folder, or both. Sonario walks all subfolders, extracts supported text, maps each document into a structured note, counts recurring patterns in Python, and synthesizes the final report.
| Lens | Intended material | Report focus |
|---|---|---|
| Auto | Mixed or unknown | Samples documents and chooses the best lens |
| Journal / Self-reflection | Journals and personal notes | Themes, emotional weight, energy, prompts |
| Work / Documentation | Project notes, meetings, specifications | Workstreams, progress, decisions, risks, open questions |
| Research / Notes | Study or research material | Concepts, findings, gaps, research questions |
| General | Anything else | Neutral themes, notable points, questions to explore |
- Lens selection: Auto samples the collection and chooses a lens.
- Extraction: Files are discovered recursively and converted to text.
- Map: Each document receives a structured analysis. Results are cached.
- Recovery: A malformed Groq JSON response gets a larger JSON retry, local JSON repair, a compact text-format recovery, and finally a local fallback. A document is not discarded solely because cloud formatting failed.
- Reduce: Python counts patterns recurring across multiple documents.
- Synthesize: The selected lens shapes the final report and follow-up questions.
A recovered item appears in progress output as:
[5/9] analyzed example.txt (JSON recovered)
Drop in a supported file or paste a link. Sonario produces Detailed, Normal, and Bullets views. Long sources are divided into sections and folded down until the result fits; a single failed section is noted rather than aborting the entire job.
| Source | Behavior |
|---|---|
| Files | .txt, .md, .rtf, .docx, .pdf, .epub, and OCR-supported images |
| YouTube | Uses available captions and provides a timestamped two-pane reader |
| Web pages | Extracts the main article text |
| Provider | Cost | Location | Notes |
|---|---|---|---|
| Qwen3.5 9B | Free | Local | Recommended default; about 6.6 GB through Ollama |
| Smart routing | Free | Local | Qwen3.5 4B for repetitive work and 9B for final synthesis |
| Qwen3.5 4B | Free | Local | Faster and lighter, with lower quality on complex sources |
| Ollama – any model | Free | Local | Enter any model already pulled into Ollama |
| Groq – Qwen 3.6 27B | Free tier, own key | Cloud | Pinned to qwen/qwen3.6-27b; text is sent to Groq |
Groq requests use conservative pacing, live reset headers, non-thinking mode, and
max_completion_tokens. Structured document mapping uses JSON Object Mode and a
layered recovery path for truncated or malformed responses.
- Local Ollama providers keep document text on the computer.
- Groq sends the text needed for the selected job to Groq's servers.
- Google Drive is read-only. Sonario does not silently authorize an account; OAuth is allowed only during an explicit connect/setup action.
- Remembered API keys are stored locally in plain text under the gitignored
credentials/directory.
Google Drive is optional. Sonario requests read-only access using a personal Google OAuth desktop client. Normal app startup checks only local credential state; it does not open an OAuth prompt or refresh a token unless the user explicitly connects or starts a Drive-backed job. Full setup steps are in BUILD.md.
.txt, .md, .rtf, .docx, .pdf, .epub, plus scanned PDFs and images via
Tesseract and Poppler OCR.
Add an OpenAI-compatible provider to models.json, restart Sonario, and it appears
in the provider dropdowns.
{
"providers": {
"lmstudio": {
"label": "LM Studio (free, local)",
"base_url": "http://localhost:1234/v1",
"model": "local-model",
"needs_key": false,
"min_interval": 0.0,
"note": "Start LM Studio's local server and load a model first."
}
}
}The defaults were tested on Windows 11 with an NVIDIA RTX 5060 Laptop GPU (8 GB VRAM), 16 GB RAM, and Python 3.12. This is a reference system, not a minimum requirement. Groq does not depend on local GPU performance.
app.py Flask server and API routes
pipeline.py mapping, reduction, synthesis, recovery, summarization
providers.py Ollama/Groq/custom provider interface and rate limiting
extract.py recursive file extraction and OCR
sources.py files, EPUB, YouTube, and web inputs
modes.py interpretation lenses
export.py Markdown and PDF export
gdrive.py optional read-only Google Drive OAuth
keystore.py opt-in local API-key storage
sysmon.py local CPU/RAM/GPU monitor
sonario_window.ps1 dedicated app window and X-close watcher
stop_sonario.ps1 process-specific manual shutdown
create_sonario_shortcut.ps1 desktop shortcut creation
sonario_launcher.vbs hidden desktop-launch bridge
static/ interface and icons
tests/ provider and recovery regression tests
BUILD.md installation, packaging, and full BAT contents
The BAT launchers shown in BUILD.md are added only to downloadable Windows release ZIPs. They must remain absent from Git history.
- No
*.batfile may be committed. .gitignoreblocks all BAT files with no exceptions.- CI fails if a tracked BAT file appears or if BUILD.md stops documenting a required launcher.
- Runtime state, logs, credentials, API keys, caches, output, browser profiles, PID files, and diagnostics are ignored.
MIT © pgotta. See LICENSE.
Ollama, local models, Groq, Google APIs, Tesseract, Poppler, YouTube, and other third-party services or tools retain their own licenses and terms.