Skip to content

Latest commit

 

History

History
225 lines (139 loc) · 3.35 KB

initial_specifications.md

File metadata and controls

225 lines (139 loc) · 3.35 KB

Product Vision

Vision

Build a deterministic, auditable, and scalable Context Layer for AI systems that allows organizations to turn documentation and structured knowledge into reliable, bounded context for local and remote agents.

The platform is designed to:

  • Reduce hallucinations
  • Enforce knowledge boundaries
  • Enable on‑prem and air‑gapped deployments
  • Integrate natively with agent systems via MCP

Target Users

  • Engineering teams deploying LLMs internally
  • Regulated industries (EU-first)
  • Platform / infra teams

Non‑Goals

  • We are not a chatbot
  • We are not a vector database replacement
  • We do not train foundation models
  • We do not optimize prompts dynamically at runtime

Open‑Core Strategy

Open Components

  • Context ingestion
  • Deterministic context selection
  • Context cache lifecycle
  • MCP server
  • CLI and SDKs

Paid Components

  • Context observability
  • Drift detection
  • Cost attribution
  • Usage analytics
  • Policy intelligence

The open core must be fully functional without paid components.

Core Concepts

Document Model

A Document is an immutable, versioned unit of knowledge.

Properties

  • id (stable)
  • version (hash-based)
  • source
  • content
  • metadata

Documents are never mutated; updates create new versions.

Context Cache

A Context Cache is a compiled, queryable snapshot of documents optimized for agent consumption.

Guarantees

  • Deterministic output
  • Explicit invalidation
  • Rebuild reproducibility

Lifecycle

  1. Ingest documents
  2. Normalize and parse
  3. Build cache
  4. Serve via MCP

Context Selection

Context selection determines which documents are included in a response.

Constraints

  • Token budget
  • Relevance score
  • Policy filters

Selection must be:

  • Deterministic
  • Explainable
  • Replayable

Summarization & Compression

Summarization is optional and offline.

Rules:

  • Never lossy by default
  • Source references must be preserved
  • Output must be traceable to original documents

MCP Interface Specification

Purpose

Expose context resolution to agents via the Model Context Protocol.

Core Methods

  • list_contexts
  • resolve_context
  • inspect_cache

resolve_context

Inputs:

  • context_id
  • query
  • budget

Outputs:

  • ordered document bundle
  • metadata
  • provenance

CLI Specification

Commands

  • ingest
  • build
  • serve
  • inspect

CLI must:

  • Be scriptable
  • Emit JSON by default
  • Support CI usage

Security Model

Principles

  • No outbound calls by default
  • Explicit trust boundaries
  • Read-only runtime

Deployment Modes

  • Local
  • On‑prem
  • Air‑gapped

Determinism Guarantees

Given:

  • Same documents
  • Same versions
  • Same configuration

The system MUST produce identical context outputs.

Observability (Paid)

Signals

  • Context hit rate
  • Cache freshness
  • Token utilization

Constraints

Observability must not alter context selection behavior.

Licensing

  • Open core licensed under Apache 2.0 or MPL
  • Paid components proprietary
  • Specs publicly readable

Architectural Decision Records

ADR‑0001: Determinism First

All context operations prioritize determinism over adaptability.

ADR‑0002: MCP as Primary Interface

MCP is the first‑class integration surface for agents.

ADR‑0003: Open Core Boundary

No paid logic may be required for correctness.