Cinestar
Privately Index and Search Through Media
4 followers
Privately Index and Search Through Media
4 followers
Search through images and videos backed by local LLMs. Private, fast, offline-first — powered by bge-large embeddings, moondream vision, and Whisper transcription. For anyone with 10,000+ photos who can never find anything
As someone who's always drowning in screenshots, videos, and random media files, I built Clipwise to solve my own problem: finding that one specific image or video moment when I actually need it.
Initial version is targeted towards:
Professional photographers/video editors who are slightly tech savy and comfortable using some ai inference tools
Tech Savy people, who have accumulated lot of photos/videos/audios over time and need a better search tool
Demo video:
Cinestar features:
🔒 100% local - Your media never leaves your machine
🧠 AI-powered - Search by describing what you see ("red car in parking lot")
⚡ Instant results - No cloud delays, everything runs on your Mac
🎯 Precise timestamps - Jump to exact moments in videos
Some of it is a work in progress, let me know the issues
# Clipwise v0.1.3
TL;DR: Search your videos, images, and audio using natural language - now 40% faster!
## 🎯 What's New
### ⚡ Performance Boost
- 40% faster search during video processing (10s → 6s)
- Smart Ollama instance routing: dedicated resources for search, load-balanced for indexing
- No more waiting - search stays responsive even during heavy video processing
### 🎬 Video Intelligence
- Videos searchable in 60 seconds after upload
- Multi-modal understanding: combines audio transcription + visual keyframes + scene context
- Integrated video player with smart segment navigation
- Search by what you see AND what you hear: "person explaining code" or "sunset beach scene"
### 🐛 Fixes
- Native dependencies now work in packaged builds (Sharp, FFmpeg)
- Logo displays correctly in splash screen
- Smoother installation experience across all platforms
## 🏗️ Architecture
- 2 Ollama instances with nginx load balancing
- Batch-concurrent processing pipeline
- SQLite + vector search for blazing fast queries
## 🎬 UI Fixes
UI BUG showing alert dialogues. Sorry FOLKS!!
---
Try it: Download, run `docker-compose up -d`, and start searching your media library with natural language!
# Clipwise v0.1.46
TL;DR: Stable Production Version
# Cinestar (name change)
Release Date: October 11, 2025
## 🎯 Overview
## ✨ Major Changes
### 1. Simplifying Installation
Removed dependency from docker, whisper is downloaded on request
Transcription is no longer availalble via api, but a cli command
### 2. Speed up local folder indexing
Local Folder connect indexing and searchability improved
Fixes in search and captioning. Allowing for tagging as well
I understand, moving/downloading another new application, could be a hassle. Overall I would still want the option to use this privately hosted. The problem is with local, one needs to install any inference server, for example Ollama , is one of the easiest.
This seems to be a possible barrier to entry, not to mention, I can monetize it better. So would it be easier to deliver this via the web and a phone app?
That way say free tier users get only 5 videos or 30 minutes, have some data already setup for demo. On the app I can provide a toggle for `Incognito` mode, with Incognito on, it only relies on local inference servers, if not, BYOK or Use ours.
I really don't want to become a video/media provider, when I can rely on external integrations, so google, dropbox is something that is planned. So I don't have much motivation to store or ask users to upload their videos separately . Infact on a phone app things become much easier.
Any opinions?
Website url changed to: https://cinestar.sourceforge.io/
New Updates Coming for:
Deduplication
Organisation
Option to use online models for people who can't run ollama and setup models
Automatic download of models without manual intervention
Internal Changes to handle data better.
This update focuses on smarter video understanding, faster performance, and a more flexible privacy model that gives you the choice between local and external AI providers.
🚀 Smarter Video Search
Video indexing is now multi-modal — combining audio transcription and visual scene understanding for precise, searchable context.
Audio Phase (Whisper): 20-min videos transcribed in ~60s
Visual Phase: Keyframe extraction + captioning for scene meaning
Smart Segmentation: Every 30s segment gets its own searchable context
Incremental Indexing: Start searching before processing finishes
It’s now easier than ever to find that exact moment in a long video — just type what was said or shown.
🖼️ Better Image Search
Images now go through multi-pass captioning — overview → details → context — giving the search engine richer understanding of each picture.
And we’ve opened up privacy settings:
You can now choose between 100% local processing via Ollama, or opt-in to external providers like OpenAI or Google Gemini for higher-quality captions and embeddings.
Drillbit stays local by default — but now you decide the balance between speed, quality, and privacy.
🎵 Audio Support
Audio files get full transcription and indexing too. Search podcasts, interviews, or voice memos by meaning — not filenames.
Waveform visualization and playback controls make navigation effortless.
🔍 Unified Search Experience
Search everything from one place — images, videos, and audio — with instant results.
Hybrid Search: Combines AI semantic matching + keyword text search
Context-Aware Results: Jump straight to relevant timestamps in videos or audio
Deduplication: Smarter grouping of segments and parent files
Real-Time Updates: Search results appear as you type
The experience feels like a private, offline version of “Google Photos + ChatGPT” — but built for your local drives.
🎨 Refreshed Interface (Driller v2)
A complete redesign focused on clarity and speed:
Sleek, glassy theme with soft animations
Scope filters (All, Folders, S3, Drive)
Activity panel to track ongoing indexing
Infinite scroll & smooth virtualization for large collections
Full-page media viewers replace old modals for a cleaner browsing flow — no z-index chaos, just focused exploration.
⚡ Performance Boosts You’ll Notice
Sub-1s total search time
~100s for full AI indexing of a 20-min video
2–3s image caption generation
0.5s embedding generation
Everything feels faster and more responsive, even while indexing is running in the background.
🔒 Privacy — Now with Options
Previously, Drillbit was strictly offline.
Now, you can selectively enable external AI providers for tasks that benefit from cloud-grade models — while keeping everything else local.
Your data is never uploaded by default. External providers are opt-in, transparent, and reversible at any time.
🧭 Coming Next
Cloud & NAS integrations (S3, Google Drive)
Smart collections & auto-clustering
Advanced filters (date, people, location)
Batch tagging & export options
PWA support
We’re focused on making Cinestar scale to 100K+ media files while keeping it private, fast, and personal.
Version: 0.1.64
Release: October 2024
License: MIT
Find anything — locally or with help from AI — while staying in control of your data.
Made with ❤️ for people who value privacy and want powerful local search