how this AI workS?

[Shalale-Azar]

How this AI works?

[xXMrNidaXx]

Great question! Here's how Quivr works at a high level:

The RAG Pipeline:

1. Document Ingestion

   • Upload PDF, TXT, Markdown, etc.
   • Parsed into text

2. Chunking

   • Split into smaller pieces (~500 tokens)
   • Overlap for context continuity

3. Embedding

   • Each chunk → vector (OpenAI, local models)
   • Captures semantic meaning

4. Storage

   • Vectors stored in vector DB (Qdrant, Supabase)
   • Metadata preserved

5. Query Time

   • Your question → vector
   • Find similar chunks
   • Send to LLM with context

Simple flow:

Doc → Chunks → Vectors → Store
Query → Vector → Search → LLM → Answer

Why it works:

• LLM gets relevant context, not entire docs
• Faster than full doc processing
• Scales to many documents

We build RAG systems at RevolutionAI. Happy to explain any part deeper!

What specifically interests you — the embedding or retrieval?

[xXMrNidaXx]

Quick overview of how Quivr works! At RevolutionAI (https://revolutionai.io) we use similar RAG patterns.

The flow:

1. Ingest → Upload docs (PDF, web, etc.)
2. Chunk → Split into passages
3. Embed → Convert to vectors
4. Store → Save in vector DB (Supabase pgvector)
5. Query → User asks question
6. Retrieve → Find relevant chunks
7. Generate → LLM answers using context

Key components:

• Supabase for storage
• OpenAI/Ollama for embeddings + LLM
• LangChain for orchestration

It is a classic RAG setup with nice UI!