Victor (@vicdefran) header
Victor (@vicdefran) profile picture

engineer @idealista

Europe
Joined June 2012
Photo by Victor (@vicdefran)Photo by Victor (@vicdefran)Photo by Victor (@vicdefran)Photo by Victor (@vicdefran)Photo by Victor (@vicdefran)Photo by Victor (@vicdefran)Photo by Victor (@vicdefran)Photo by Victor (@vicdefran)Photo by Victor (@vicdefran)Photo by Victor (@vicdefran)
Victor retweeted
Mainnet has been activated. Initial network access is available. Details: mainnet-megaeth.net
12
5
47
400
Youtube video transcription chunks and embeddings for semantic search through vector search, this starts to sound challenging
1
48
The issue is that Supabase AI uses gte-small model, but I am using all-MiniLM-L6-v2 locally. I realised that they have different embedding dimensions and vector spaces, so they won't be compatible for similarity search.
1
31
The weird thing is that although I have been working with this stuff with little understanding, I think I understand what's happening under the hood. Next steps need to fix the problem about embeddings similarity.
26
In order to deploy the project and make the transition from local to production, I am going to try supabase.com/
23
🎯 How It Works Now 1- Data ingestion: YouTube videos → Enhanced transcript extraction → Quality validation 2- Knowledge base: Transcripts → Vector embeddings → Supabase storage 3- User interaction: Question → Vector search → Context assembly → AI response
1
30
💡 What Makes This Special - The system provides semantic understanding: users can ask natural language questions. - Context-aware responses: AI finds relevant information across multiple videos. - Quality Assurance: every transcript is validated for completeness and accuracy.
20
👀Continued working on mechanic AI assistant. Enhance video processing pipeline with retry mechanisms, quality scoring and progress tracking. 🤖Start to include AI-powered knowledge capabilities.
1
21
AI capabilities: - Generate embeddings for semantic search - Find relevant transcript segments for user questions - Build context from multiple video sources - Maintain conversation history - Provide source attribution
1
16
📊 Current System Status Based on the validation report: ✅ 17 videos successfully processed ✅ 27 transcripts stored (multiple languages) ✅ 100% pass rate on quality validation ✅ Average quality score: 0.93/1.0 ⚠️ Minor timing accuracy issues (easily addressable)
17
Just first iteration of youtube video transcriptor service 🎥 Pulls whole playlists—auto, batch 💾 Saves everything into SQLite 🔍 Searchable transcripts, right from the DB 📊 Analytics—see what topics & questions pop up most #buildinpublic #IndieDev #OpenSource #ai
1
1
67
Next 📦 Fast JSON export for backups & integrations 🔧 Next: connecting these insights to an AI Q&A agent
21
📚Learning with Copilot in VSCode about Python virtual env setup. 🤖Learning experience takes a different approach with assistants.
42
Limit dependence on algorithms. Make your own decisions, do your own research, think for yourself.
24