SourceLens

KO | EN

gitlite — search

#javascript #python #typescript #llm #machine-learning #rust #linux #docker #windows #vue #mcp #ios #openai #open-source

(★ 24)

Multilingual source-grounded research workspace with agentic RAG, cited chat, podcasts, PowerPoint, and video.

SourceLens 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<div align="center">
  <img src="src/paperlens/web/favicon.svg" width="88" alt="SourceLens logo">
  <h1>SourceLens</h1>
  <p><strong>Multilingual research, grounded in evidence.</strong></p>
  <p>
    Build a cited workspace from documents, media, notes, and the web—then turn that evidence<br>
    into conversations, podcasts, editable slides, videos, reports, study aids, and visual summaries.
  </p>

[![CI](https://github.com/msmahdinejad/SourceLens/actions/workflows/ci.yml/badge.svg)](https://github.com/msmahdinejad/SourceLens/actions/workflows/ci.yml)
  [![Release](https://img.shields.io/badge/release-v2.2.2-0f766e)](https://github.com/msmahdinejad/SourceLens/releases/tag/v2.2.2)
  [![Python 3.12](https://img.shields.io/badge/Python-3.12-3776AB?logo=python&logoColor=white)](https://www.python.org/)
  [![Node 20+](https://img.shields.io/badge/Node.js-20%2B-339933?logo=nodedotjs&logoColor=white)](https://nodejs.org/)
  [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://raw.githubusercontent.com/msmahdinejad/SourceLens/main/LICENSE)
</div>

![SourceLens workspace with processed research sources](https://raw.githubusercontent.com/msmahdinejad/SourceLens/main/docs/demo/assets/02-sources-ready-en.png)

SourceLens is a self-hosted research workspace for people who want the convenience of an AI
notebook without losing sight of the evidence. It combines page-aware extraction, selective OCR
and vision, agentic retrieval, inspectable citations, durable job events, and professional local
rendering in one application.

The installable Python package and CLI are named `paperlens` for compatibility.

## Why SourceLens

| Capability | What you get |
| --- | --- |
| Grounded conversation | Source selection, page and segment citations, an evidence drawer, and explicit abstention when the workspace cannot support an answer |
| Agentic RAG | The model can plan multiple searches, choose filters and evidence budgets, and repair invalid tool calls without crashing the request |
| Hybrid retrieval | Chroma vector search and SQLite FTS5 fused for semantic matches, exact terms, identifiers, equations, and Persian text |
| Multimodal ingestion | PDF, text, images, audio, video audio tracks, pasted notes, prompt images, and selected web pages |
| Adaptive extraction | Native text first; OCR and visual analysis only when document signals justify the extra work |
| Research Studio | 13 source-scoped workflows for audio, PowerPoint, video, reports, maps, cards, quizzes, charts, comparisons, and more |
| Live progress | Replayable events, accurate progress, cancellation, retries, cooldowns, and restart recovery |
| Multilingual UX | English LTR and Persian RTL interfaces, retrieval, chat, and Studio outputs |
| Local ownership | Workspaces, sources, indexes, conversations, and generated artifacts stay on the host machine |

## A real bilingual demo

The **Charge While You Shop** walkthrough asks a practical question: can retail car parks become
convenient, reliable, interoperable, and grid-aware EV charging stops? SourceLens processed two
openly licensed studies, added five curated web sources through Search, indexed **236 chunks**, and
generated every Studio format in both English and Persian—**26 application-generated outputs** in
total. The refreshed overview videos run for **50 seconds**, with exact runtimes recorded
in their public manifests.

**[Explore the English demo and downloads](docs/demo/README.md)** ·
**[مشاهدهٔ دمو و خروجی‌های فارسی](docs/demo/README.fa.md)**

Evaluation reports: [English PDF](docs/demo/outputs/SourceLens-demo-evaluation-en.pdf) ·
[English PowerPoint](docs/demo/outputs/SourceLens-demo-evaluation-en.pptx) ·
[Persian PDF](docs/demo/outputs/SourceLens-demo-evaluation-fa.pdf) ·
[Persian PowerPoint](docs/demo/outputs/SourceLens-demo-evaluation-fa.pptx)

### Bring sources into one workspace

![A PDF moving through the SourceLens ingestion pipeline](https://raw.githubusercontent.com/msmahdinejad/SourceLens/main/docs/demo/assets/01-source-ingestion-en.png)

Extraction, OCR, visual analysis, transcription, chunking, and indexing publish their progress as
they happen. A source remains usable when optional enrichment is unavailable.

### Search before you import

![Reviewable web search results in SourceLens](https://raw.githubusercontent.com/msmahdinejad/SourceLens/main/docs/demo/assets/04-web-search-en.png)

Web results stay reviewable before import. Selected pages retain their URL and publisher metadata
and enter the same workspace boundary as uploaded files.

### Turn evidence into useful formats

![A source-grounded concept map in Studio](https://raw.githubusercontent.com/msmahdinejad/SourceLens/main/docs/demo/assets/06-concept-map-en.png)

![Source-grounded flashcards in Studio](https://raw.githubusercontent.com/msmahdinejad/SourceLens/main/docs/demo/assets/08-flashcards-en.png)

![A structured evidence report in Studio](https://raw.githubusercontent.com/msmahdinejad/SourceLens/main/docs/demo/assets/10-report-en.png)

Claims remain tied to the selected sources, while each completed artifact can be inspected and
downloaded from its Studio card.

### Produce editable slides, audio, and video

![Completed PowerPoint artifact in SourceLens Studio](https://raw.githubusercontent.com/msmahdinejad/SourceLens/main/docs/demo/assets/13-slide-deck-studio-en.png)

![Completed video overview in SourceLens Studio](https://raw.githubusercontent.com/msmahdinejad/SourceLens/main/docs/demo/assets/15-video-studio-en.png)

The visual renderers preserve citations, typography, safe margins, captions, notes, and media
provenance. Video narration stays at its natural speed.

### Work naturally in Persian

The complete interface, retrieval flow, cited chat, evidence panel, and Studio output controls
support Persian RTL as well as English LTR. See the [Persian demo page](docs/demo/README.fa.md) and
[Persian evaluation report](docs/demo/outputs/SourceLens-demo-evaluation-fa.pdf).

## Architecture

```mermaid
flowchart LR
    A["Files, notes, media, web pages"] --> B["Adaptive extraction"]
    B --> C["Page-aware chunks"]
    C --> D["Chroma vectors"]
    C --> E["SQLite FTS5"]
    D --> F["Rank fusion + evidence balancing"]
    E --> F
    Q["Question + selected sources"] --> G["Agentic tool loop"]
    F --> G
    G --> H["Cited answer + evidence"]
    G --> I["Studio plan"]
    I --> J["Audio / PPTX / MP4 / structured outputs"]
    I --> K["SQLite jobs + append-only events"]
    K --> L["SSE live progress"]
```

SourceLens is a modular monolith built with FastAPI, native ES modules, SQLite WAL/FTS5, Chroma,
bounded workers, Node renderers, and FFmpeg. The local single-node product does not require Redis,
Celery, Postgres, a frontend build chain, or a graph database.

## Quick start

### Requirements

- Python 3.12
- [uv](https://docs.astral.sh/uv/)
- Node.js 20 or newer
- FFmpeg for video composition and optional MP3 export

### Windows

```powershell
git clone https://github.com/msmahdinejad/SourceLens.git
cd SourceLens
uv sync
npm ci --prefix src/paperlens/renderers/deck_engine
npm ci --prefix src/paperlens/renderers/video_engine
Copy-Item .env.example .env
uv run paperlens web
```

### Linux / macOS

```bash
git clone https://github.com/msmahdinejad/SourceLens.git
cd SourceLens
uv sync
npm ci --prefix src/paperlens/renderers/deck_engine
npm ci --prefix src/paperlens/renderers/video_engine
cp .env.example .env
uv run paperlens web
```

Open [http://127.0.0.1:8000](http://127.0.0.1:8000). A new installation starts with one empty
local workspace.

Optional local capabilities:

```powershell
uv sync --extra ocr    # Pillow + Tesseract adapter; install Tesseract separately
uv sync --extra voice  # Faster-Whisper transcription fallback
```

## Providers

Copy `.env.example` to `.env` and configure only the services you want. SourceLens includes routes
for Gemini, Groq, Mistral, NVIDIA NIM, OCR.Space, Exa, Tavily, Pexels, and Jamendo, plus custom
OpenAI-compatible endpoints.

```dotenv
GOOGLE_API_KEYS=google_key_1,google_key_2
GROQ_CHAT_API_KEYS=groq_key_1
MISTRAL_API_KEYS=mistral_key_1
NVIDIA_API_KEYS=nvidia_key_1

EXA_API_KEYS=exa_key_1
TAVILY_API_KEYS=tavily_key_1
PEXELS_API_KEYS=pexels_key_1
JAMENDO_CLIENT_IDS=jamendo_client_id
```

Provider and model priority, capabilities, cooldowns, pacing, OCR/VLM limits, and retrieval settings
can also be managed in the app. See the [provider model catalog](docs/provider-model-catalog.md).

## Retrieval and citations

The research assistant exposes a small set of allowed tools to the model. It can run focused
multi-search plans, choose document/page/section filters, balance evidence across sources, and make
follow-up calls before synthesis. Every final document claim must resolve to selected evidence;
missing context follows an explicit insufficient-evidence path.

The versioned benchmark currently covers 13 pinned CC0 Wikidata revisions, 52 English/Persian
chunks, and 45 questions. FTS5/vector fusion measured **100% Recall@6**, **79.6% citation
precision**, and **100% citation recall**. Alias, multi-hop, and contradiction recall also reached
**100%**, keeping the predefined knowledge-graph gate closed. Results and the graph decision are
documented in the [retrieval benchmark](docs/retrieval-benchmark.md).

## Studio outputs

- **Audio overview** — narrated, scripted, producer-guided, or autonomous; one to five speakers;
  reviewed scripts, turn checkpoints, natural speed, audio, and transcript.
- **Slide deck** — editable PowerPoint, RTL-aware layouts, formulas, charts, verified media,
  citations, speaker notes, and presenter/organization metadata.
- **Video overview** — autonomous scene direction, verified media, natural narration, captions,
  transitions, exact duration, credits, timeline, and diagnostics.
- **Structured research** — reports, concept maps, flashcards, quizzes, infographics, tables,
  charts, timelines, source comparisons, and research-gap analysis.

## Reliability and privacy

- Workspace-local SQLite WAL stores jobs, attempts, leases, revisions, and append-only events.
- SSE replays progress after reconnect; bounded polling is a fallback.
- Cancellation propagates through provider waits, downloads, Node, and FFmpeg processes.
- Startup recovery resumes checkpoint-safe work and marks unsafe work retryable.
- Uploaded names never become storage paths; generated IDs preserve workspace isolation.
- Credentials and workspace content are redacted from status, traces, errors, and release artifacts.

Read [SECURITY.md](SECURITY.md) before exposing the trusted local server to a network. Authentication,
tenant authorization, malware scanning, HTTPS, and a production task queue remain deployment
responsibilities rather than hidden guarantees of the local edition.

## Optional demo corpus

```powershell
uv run paperlens demo
```

The command downloads two CC BY 4.0 electric-mobility studies from their publishers, verifies
pinned SHA-256 checksums, and indexes them through the same ingestion pipeline as user documents.
The PDFs are not stored in Git or distributed in packages. See the
[source and rights note](docs/demo/source-research.md) and
[evidence-backed prompt set](docs/demo_questions.md).

## Development

```powershell
uv sync --extra dev
uv run ruff check .
uv run pytest
uv run pytest --cov=paperlens --cov-report=term-missing
node --check src/paperlens/web/app.js
node --test tests/node/video_engine.test.mjs tests/node/studio_live.test.mjs
npm audit --omit=dev --prefix src/paperlens/renderers/deck_engine
npm audit --omit=dev --prefix src/paperlens/renderers/video_engine
uv run python scripts/license_audit.py
python scripts/repository_policy.py audit
uv build
```

Continuous integration runs the Python and Node suites, dependency and license audits, public-tree
checks, and reproducible package builds. Read [CONTRIBUTING.md](CONTRIBUTING.md) before opening a
pull request.

## Documentation

- [Architecture](docs/architecture.md)
- [Operations, backup, and recovery](docs/operations.md)
- [Provider model catalog](docs/provider-model-catalog.md)
- [Retrieval benchmark](docs/retrieval-benchmark.md)
- [Visual output architecture](docs/visual-output-architecture.md)
- [Charge While You Shop demo](docs/demo/README.md)
- [Demo source and rights note](docs/demo/source-research.md)
- [Security policy](SECURITY.md)
- [Contributing guide](CONTRIBUTING.md)

## License

SourceLens-authored source code and original project assets are licensed under the
[MIT License](LICENSE). Dependencies and bundled assets retain their own licenses.

PDF extraction uses permissively licensed pdfplumber/pdfminer.six; rendering uses pypdfium2 and
PDFium with the required notices. Vazirmatn is distributed under SIL OFL 1.1. See
[THIRD_PARTY_LICENSES.md](THIRD_PARTY_LICENSES.md) for the full dependency and asset notice.

SourceLens

SourceLens 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

SourceLens

SourceLens 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README