Backend

Vectora’s backend infrastructure is designed to solve the biggest bottleneck for modern AI agents: state management at scale. While the Vectora runtime processes intelligence, the backend ensures that this intelligence is grounded in persistent, secure, and searchable data.

Vectora uses a unified architecture based on MongoDB Atlas, allowing vectors, metadata, and operational memory to coexist within the same ecosystem.

The Engine Behind Context

Topics in this Section

Page	Description
MongoDB Atlas	Why we chose Atlas and how it serves as our data foundation.
Vector Search	Technical deep dive into embeddings, cosine similarity, and the HNSW algorithm.
Persistence and Memory	How Vectora maintains state between sessions and builds long-term memory (`AGENTS.md`).

Architecture Overview

Vectora’s backend is not just a storage location; it is an extension of the Harness Runtime.

    graph TD
    A[Vectora Runtime] --> B{Service Layer}
    B --> C[Vector Service]
    B --> D[Session Service]
    B --> E[Audit Service]

    subgraph "MongoDB Atlas (Managed by Kaffyn)"
        C --> F[(Documents & Vectors)]
        D --> G[(Operational State)]
        E --> H[(Audit Logs)]
    end

Backend Principles

Namespace Isolation (RBAC): The backend enforces strict boundaries. Data from one project never mixes with others, ensuring multi-tenant privacy.
Atomicity: Code vectors and metadata are stored together. If a file is updated, the vector index and the document are updated simultaneously.
Transparent Scalability: Using MongoDB Atlas, Vectora scales from small individual projects to enterprise codebases with millions of lines of code without performance loss.
Active Governance: Every interaction is persisted and auditable, allowing full transparency over AI actions.

Backend Modes

Mode	Description	Ideal Use
Managed (SaaS)	MongoDB Atlas backend provisioned by Kaffyn. Zero configuration.	Free, Pro, and Team plans.
Hybrid (BYOK)	You provide Voyage API keys, but Atlas is managed by Kaffyn.	API cost control with easy infrastructure.
Enterprise / Custom	Connection to your own MongoDB Atlas cluster or on-premise infrastructure.	Strict data sovereignty requirements.

Frequently Asked Questions

Q: Where are my vectors physically stored? R: In MongoDB Atlas clusters managed by Kaffyn, typically located in low-latency AWS or Google Cloud regions.

Q: Is my backend data used to train Kaffyn’s models? R: No. We follow a strict privacy policy. Your vectors and metadata are your exclusive property and used only to provide context to your agent.

Q: Is the backend required for offline mode? R: Vectora allows local caching of some metadata, but full semantic vector search and long-term persistent memory depend on backend connectivity.

External Linking

Concept	Resource	Link
MongoDB Atlas	Atlas Vector Search Documentation	www.mongodb.com/docs/atlas/atlas-vector-search/
MCP	Model Context Protocol Specification	modelcontextprotocol.io/specification
MCP Go SDK	Go SDK for MCP (mark3labs)	github.com/mark3labs/mcp-go
RBAC	NIST Role-Based Access Control Standard	csrc.nist.gov/projects/rbac
RAG	Retrieval-Augmented Generation for Knowledge-Intensive NLP	arxiv.org/abs/2005.11401
HNSW	Efficient and robust approximate nearest neighbor search	arxiv.org/abs/1603.09320

Phrase to remember: “The runtime is the brain; the backend is the library. Without an organized library, the brain has nowhere to look for answers.”

Part of the Vectora ecosystem · Open Source (MIT) · Contributors