Project

Recall-OS

RecallOS is an AI-native enterprise knowledge operating system that allows organizations to ingest, organize, search and reason over every piece of company knowledge.

AI-native knowledge operating system for multimodal enterprise search and long-term organizational memory.

RecallOS ingests documents, images, videos, audio and conversations into a unified knowledge base that supports semantic search, hybrid retrieval, memory, and agentic workflows.

Features

  • 📄 Multimodal ingestion (PDFs, PPTs, Images, Audio, Video)
  • 🔍 Hybrid retrieval (SPLADE + dense vector search)
  • 🧠 Long-term organizational memory
  • 🤖 Agent-powered search and reasoning
  • 📚 Source-grounded responses with citations
  • ⚡ Distributed ingestion pipeline

Tech Stack

Layer Technology
Frontend Next.js
Backend Express
Workers Redis Streams
Storage MinIO (S3)
Database PostgreSQL
Vector DB Qdrant
Sparse Retrieval SPLADE
Parsing LlamaParse
Audio Whisper + CLAP
Vision Vision-Language Models
LLM Provider Agnostic

Architecture

User Upload
      │
      ▼
   MinIO (S3)
      │
      ▼
 Redis Streams
      │
      ▼
Distributed Workers
      │
      ▼
 LlamaParse / Whisper / Vision
      │
      ▼
 Semantic Chunking
      │
      ▼
 Chunk Enrichment
      │
      ▼
Dense Embeddings + SPLADE
      │
 ┌────┴─────┐
 ▼          ▼
Qdrant   Sparse Index
      │
      ▼
 PostgreSQL Metadata

Retrieval Pipeline

User Query
      │
      ▼
Query Rewriting
      │
 ┌────┴─────┐
 ▼          ▼
Dense Search    SPLADE Search
(Qdrant)        (Sparse)
      │          │
      └────┬─────┘
           ▼
    Reciprocal Rank Fusion
           ▼
      Cross-Encoder
        Reranker
           ▼
        Context
           ▼
           LLM

Multimodal Retrieval

Documents

  1. Semantic chunking
  2. Context enrichment
  3. Dense + SPLADE indexing

Images

  1. Vision-language embeddings
  2. OCR + metadata
  3. Text ↔ Image retrieval in a shared embedding space

Audio

  1. Whisper transcription
  2. CLAP embeddings for shared audio-text retrieval

Video

  1. Transcript indexing
  2. Keyframe extraction
  3. Vision-language embeddings

Principles

  1. Memory-first
  2. Source-grounded
  3. Agent-native
  4. Horizontally scalable
  5. Multimodal by design