Dolos Secure AI Workspace by Forensic Intelligence
Secure initialization Systems online

Dolos

Getting the software ready

Decrypting your private workspace, syncing secure modules, and preparing zero-trust chat.

Initializing secure session...

Private inference • Encrypted at rest
Back to How It Works

Message Flow

Trace exactly how your message travels through our system — encrypted, classified, and processed.

Real-Time Streaming
End-to-End Encrypted
Intelligent Routing

Choose a Scenario

Scroll to follow the message

  1. User Input

    You type your message in the chat interface.

  2. E2EE Encryption

    Your browser negotiates a per-session AES-256-GCM key at connect time (via RSA-OAEP-256, 3072-bit). Every outgoing message is encrypted with this key before transmission. The key exchange happens once; all subsequent encryptions are sub-millisecond AES operations.

    Encryption happens entirely in your browser — plaintext never leaves your device.

    RSA-OAEP-256 + AES-256-GCM

  3. Secure Transmission

    The encrypted payload is transmitted over a Socket.IO WebSocket connection secured by TLS 1.3.

    TLS 1.3

  4. Server Decryption

    The server's RSA private key decrypts the ephemeral AES key, then AES-256-GCM decrypts the payload in a secure processing endpoint.

  5. Semantic Classification

    Your message is classified by an embedding-based semantic router — no keyword matching, pure understanding.

    Conversation
    Web research
    Documents
    Files & automation
    Memory

    Illustrative example — not exact system routing

  6. LLM Context Assembly

    The complete conversation context is assembled as a structured message array for the AI model.

    • System prompts (identity, language, persona)
    • Tool instructions and capabilities
    • Conversation history (from vector store)
    • Pre-dispatch tool results
    • Current user message

  7. Adaptive AI Core

    Adaptive compute — quick answers stay fast, complex requests get the full engine. Native tool calling, bounded reasoning rounds.

  8. Tool Execution Loop

    The AI model requests additional tool calls. Each tool executes, its result is injected, and the model is re-invoked to synthesize.

    LLM requests tool
    Tool executes
    LLM re-invoked
    Round 1 → 2

  9. Hybrid Retrieval

    A dedicated sub-station for document-heavy requests — your question is refined into variants, parallel semantic and keyword legs are fused, reranked, budgeted, and expanded to full sections before reaching the AI core.

  10. Response Synthesis

    The AI composes the final response, incorporating all tool results, search findings, and generated files.

  11. Encrypted Response Streaming

    AI response chunks, tool results, and all server events are individually encrypted with your session's AES-256-GCM key and streamed back in real time via WebSocket. Every chunk is a separate encrypted envelope — even streaming responses are never plaintext on the wire.

    AES-256-GCM Answer verified against sources

  12. Encrypted Persistence

    Your message and the AI's response are stored with field-level encryption.

    Field-level AES-256-GCM at rest, plus an encrypted semantic index for later retrieval.

    AES-256-GCM • PBKDF2-HMAC-SHA256 Memory quietly updated
Max 3 agents per chat

Agent Deployment

An autonomous agent is deployed to handle your complex request with multi-step planning and execution.

Operation modes: Approval (user confirms) or Automatic (immediate execution)

Clarification Phase

The agent analyzes your request and generates structured clarification questions in your language to refine understanding.

Mandatory for root agents

Planning Phase

The agent creates a detailed execution plan (PLAN_JSON) with ordered tool steps, expected outputs, and sub-agent assignments.

PLAN_JSON Ordered tool_flow with purpose and expected output per step
PLAN_JSON — Max 24 steps

Execution Phase

The root agent follows the plan step-by-step, spawning parallel sub-agents for independent tasks. Each sub-agent has its own tool-calling loop.

Root Agent
Results Collected
  • Up to 3 sub-agents in parallel
  • 10-minute timeout per sub-agent
  • Sub-agents cannot spawn further sub-agents
Max 3 sub-agents — 10min each

File Delivery Guard

If the task requested output files (Excel, Word, PDF), the system verifies they were actually generated and re-runs generation if needed.

Enforced for file-output tasks

Agent Synthesis

The parent agent compiles all sub-agent results, attaches generated files, and produces a comprehensive final response.

Three Encryption Layers

Your data is protected at every stage — in transit, at the application layer, and at rest in the database.

Transport Layer (TLS 1.3)

All API traffic and WebSocket connections are encrypted with TLS 1.3, preventing interception and man-in-the-middle attacks.

TLS 1.3

E2EE Payload (RSA-OAEP + AES-256-GCM)

All communication is encrypted with AES-256-GCM session keys negotiated via RSA-OAEP (3072-bit) at connection time. Both request and response payloads — including real-time streaming, API responses, and file transfers — are encrypted at the application layer, above TLS, so no network intermediary ever sees plaintext.

RSA-OAEP-256 + AES-256-GCM

Field-Level Database (AES-256-GCM)

Sensitive database columns — messages, system prompts, access tokens — are individually encrypted with AES-256-GCM before persistence. Key derivation uses PBKDF2-HMAC-SHA256 with 600,000 iterations.

AES-256-GCM + PBKDF2

Complete Tool Registry

All tools available to the semantic router and AI model — organized by category.

Search & Research

Internet Search

Anonymized web search with source summaries and citation links.

Internet mode
Deep Research

Comprehensive multi-source analysis with expert-level insights and cited sources.

Internet mode
Fetch Website

Extract and process content from any web page. Based on modified open-source scraping software — please verify that web scraping is a legal activity in your jurisdiction.

Internet mode

Documents & Files

Document Search

Semantic search across your uploaded documents with AI-powered reranking.

Runs locally
Create File

Generate Word, Excel, PDF, and Markdown files from conversation context.

Runs locally
Edit Document

Open and visually edit documents and tables in the built-in rich text editor.

Runs locally
Compress File

Create ZIP archives from your AI Storage files.

Runs locally
Translation

Translate documents and text to any supported language.

Runs locally

Media Analysis

Image Analysis

Analyze images using advanced vision AI — describe scenes, identify objects, answer questions.

Runs locally
Video Analysis

Scene-by-scene video analysis with frame extraction and description.

Runs locally
Extract Structured Data

Extract tables, forms, and text from images beyond traditional OCR capabilities.

Runs locally
Audio Transcription

Transcribe audio files and video audio tracks in 8+ languages, fully local.

Runs locally

Location & OSINT

GeoInspector

Determine where a photo was taken using topographical AI analysis.

Runs locally
Device Location

Access the user's current device coordinates via browser geolocation.

Runs locally
Spector OSINT Agent

Guided Spector investigations: manage cases, inspect entities, choose compatible providers, and run temporary or case-linked OSINT searches.

Runs locally

Development

Skill Agent

Advanced file generation agent for complex document creation tasks.

Sandboxed
Deploy Agent

Autonomous multi-step agent with sub-agent orchestration for complex tasks.

Runs locally

Memory

Store Memory

Persist preferences and instructions across conversations.

Runs locally
Search Memory

Search stored memories by semantic similarity.

Runs locally

Utilities

Date & Time

Get the current date and time in the user's configured timezone.

Runs locally
Download Link

Generate direct download links for URLs, videos, and media.

Internet mode
File Management

List and read files in your AI Storage directory.

Runs locally