Welcome to Cogniverse¶

v0.1.0 · Multi-Agent AI Platform

Cogniverse is a self-optimizing multi-agent platform for intelligent processing and search across multi-modal content (video, audio, images, documents). It features A2A agent orchestration, continuous learning via DSPy, and complete multi-tenant isolation.

Platform Overview¶

flowchart TD
    %% Main Request Flow
    user(("<span style='color:#000'>User</span>")) --> |request| agentlayer["<span style='color:#000'><b>Agents (A2A)</b><br/>Router · Orchestrator · Search · ...</span>"]
    agentlayer --> |response| user
    agentlayer <--> store[("<span style='color:#000'>Content Store</span>")]
    agentlayer <--> |context| memory["<span style='color:#000'>Memory</span>"]
    agentlayer -.-> model["<span style='color:#000'>Model</span>"]
    config["<span style='color:#000'><b>Config</b><br/>Models · Embeddings</span>"] -.-> model
    config -.-> store

    %% Content Ingestion
    content[/"<span style='color:#000'>Multi-Modal Content</span>"/] --> |ingest| store

    %% Continuous Improvement
    agentlayer -.-> |traces| telemetry["<span style='color:#000'>Telemetry</span>"]
    telemetry -.-> evaluator["<span style='color:#000'>Evaluator</span>"]
    evaluator -.-> optimizer["<span style='color:#000'>Optimizer</span>"]
    optimizer -.-> |improves| agentlayer

    %% Experiments
    telemetry -.-> experiments["<span style='color:#000'>Experiments</span>"]
    experiments -.-> evaluator

    %% Training
    store -.-> |samples| generator["<span style='color:#000'>Synthetic Generator</span>"]
    generator -.-> trainer["<span style='color:#000'>Trainer</span>"]
    evaluator -.-> |annotations| trainer
    trainer -.-> |adapters| model

    style user fill:#90caf9,stroke:#1565c0,color:#000
    style content fill:#a5d6a7,stroke:#388e3c,color:#000
    style agentlayer fill:#ce93d8,stroke:#7b1fa2,color:#000
    style model fill:#81d4fa,stroke:#0288d1,color:#000
    style store fill:#90caf9,stroke:#1565c0,color:#000
    style memory fill:#90caf9,stroke:#1565c0,color:#000

    style telemetry fill:#a5d6a7,stroke:#388e3c,color:#000
    style experiments fill:#a5d6a7,stroke:#388e3c,color:#000
    style evaluator fill:#a5d6a7,stroke:#388e3c,color:#000
    style optimizer fill:#ffcc80,stroke:#ef6c00,color:#000
    style generator fill:#ffcc80,stroke:#ef6c00,color:#000
    style trainer fill:#ffcc80,stroke:#ef6c00,color:#000
    style config fill:#b0bec5,stroke:#546e7a,color:#000

Key Features¶

• Multi-Modal Processing — Video, audio, images, documents with unified embeddings
• Agent Orchestration — A2A protocol with OrchestratorAgent entry point, agent discovery via AgentRegistry
• 23 Specialized Agents — Gateway/routing (GatewayAgent, OrchestratorAgent, ProfileSelection, QueryEnhancement, EntityExtraction), search & analysis (Search, ImageSearch, Document, TextAnalysis, AudioAnalysis), generation (Summarizer, DetailedReport), research & coding (DeepResearch, Coding), and 9 knowledge/audit agents — see Agents Module for the full roster
• Knowledge Management Layer — Schema-driven memory with provenance tracking, contradiction detection, trust ranking, federation (org trunk + tenant overlays), pinning, and schema-driven lifecycle cleanup
• 9 Knowledge Agents — MultiDocumentSynthesis, KnowledgeGraphTraversal, CrossTenantComparison, ContradictionReconciliation, CitationTracing, TemporalReasoning, FederatedQuery, KnowledgeSummarization, AuditExplanation
• Self-Optimization — Continuous learning via GEPA, MIPRO, and synthetic data; regression-reject gate; per-tenant canary promotion; rollback
• LLM Fine-Tuning — End-to-end pipeline from telemetry to LoRA adapters
• Memory — Schema-governed context persistence with trust-ranked retrieval and per-schema lifecycle policies
• Experiments — Run and track evaluation experiments with datasets
• Hybrid Search — BM25 + dense vectors with multiple ranking strategies
• Sandbox Isolation — OpenShell per-agent sandboxing with SandboxPolicy (required/optional/disabled) and gateway health probes
• Configurable — Pluggable models, embeddings, and backends
• Multi-Tenant — Schema-per-tenant isolation with independent telemetry
• Production Ready — OpenTelemetry tracing, metrics, and distributed traces

Quick Start¶

# Install
git clone <repository-url> && cd cogniverse
uv sync

# Start services (Vespa, Phoenix, Ollama)
cogniverse up

# Open the web client
open http://localhost:28400

See Getting Started for ingestion and configuration.

Built With¶

Documentation¶