Skip to content

Repository files navigation

Knowledge Distillery

A system that delivers only verified knowledge to AI coding agents.

English | 한국어


The Problem

When AI coding agents receive too much unfiltered context, they don't get smarter — they get noisier. Unverified hypotheses are treated as facts, revoked decisions persist as instructions, and irrelevant information buries the critical rules. The result: hallucinations, silent mistakes, and costly rework.

The Approach

Knowledge Distillery takes a different path: collect broadly, distill rigorously, deliver minimally.

A 3-layer architecture with a convention-based air gap separates raw information from verified knowledge:

  1. History Archive — Raw data preserved without loss (Slack, Linear, PR reviews, AI session transcripts)
  2. Distillation Pipeline — AI-driven refinement with automated quality gates, extracting only confirmed decisions and validated anti-patterns
  3. Knowledge Vault — A curated SQLite store of distilled insights, accessible only through the knowledge-gate CLI

The agent never touches raw data. It only sees what has survived distillation.

Key Design Decisions

  • Convention-based air gap: Operational isolation between raw data and the agent's workflow — not perfect enforcement, but sufficient and pragmatic
  • AI-autonomous distillation: LLMs extract and compress; humans provide strategic oversight, not item-by-item approval
  • Context Gate: Domain-based selective filtering ensures agents receive only task-relevant knowledge, not the entire vault
  • Append-only, two types only: The vault contains only Facts and Anti-Patterns — no confidence scores, no hedging

Documentation

Executable skills in plugins/knowledge-distillery/skills/ are the source of truth for pipeline behavior.

Document Description
Design Philosophy Why this architecture exists — the rationale behind information control for AI agents
Implementation Design How it works — pipeline design, vault schema, runtime policies, deployment as a Claude Code Plugin
Adoption Guide How to install the plugin, initialize a repository, and start using the vault
Tool Evaluation Adopted and rejected tools with rationale

CLI Quick Reference

The knowledge-gate CLI is the sole access path to the knowledge vault (.knowledge/vault.db). Set KNOWLEDGE_VAULT_PATH to override the default vault location.

Agent Runtime — query rules before modifying code:

Command Purpose
query-paths [--ids-only] <filepath> Resolve file path to domains → return a lightweight summary index, or IDs only
query-domain [--ids-only] <domain> Query rules by domain name → return a lightweight summary index, or IDs only
search [--ids-only] <keyword> FTS5 full-text keyword search → return a lightweight summary index, or IDs only
get <id> Retrieve full entry details (including body)
get-many <id>... Retrieve full details for multiple entries in one call
list [--ids-only] Summary list of all active entries, or IDs only

Domain — explore and manage the domain registry:

Command Purpose
domain-info <domain> Domain details (description, patterns, entry count)
domain-list [--status X] [--ids-only] List domains or emit a lightweight domain ID index
domain-resolve-path <filepath> Reverse-lookup file path to domains
domain-add, domain-merge, domain-split, domain-deprecate Registry lifecycle
domain-paths-set, domain-paths-add, domain-paths-remove Path pattern management
domain-report Domain health diagnosis

Loading — add entries to the vault:

Command Purpose
add --type <type> --title ... --claim ... --body ... --domain ... [flags] Add entry with validation
_pipeline-insert Pipeline-only bulk INSERT (JSON stdin)

Utility — setup and maintenance:

Command Purpose
init-db [path] Create a new vault from bundled schema
migrate Apply schema migrations
curate Interactive curation queue resolution

Examples:

# Query rules before editing a file
knowledge-gate query-paths src/api/auth/login.ts

# Query a domain with an ultra-light ID-only index
knowledge-gate query-domain --ids-only payment

# Fetch multiple full entries after a lightweight discovery step
knowledge-gate get-many global-convention service-rule

# Check which domains a file belongs to
knowledge-gate domain-resolve-path src/services/payment.rb

# Preload a lightweight domain ID index for navigation
knowledge-gate domain-list --ids-only

# Search for rules by keyword
knowledge-gate search "callback"

Run knowledge-gate help for full usage details.

Delivery Format

Knowledge Distillery supports two delivery paths:

  • Portable agent skills: npx skills add https://github.com/ether-moon/knowledge-distillery/tree/main/plugins/knowledge-distillery --skill '*' -a codex
  • Claude Code Plugin: install with claude plugin install to use the /knowledge-distillery:* namespace

Each skill owns the scripts and assets it needs. The knowledge-gate CLI and schema live inside the knowledge-gate skill; hook templates and workflow assets live inside the setup skill. Running setup installs lifecycle hooks into the active host's repo-local configuration. The runtime remains vendor-neutral and uses sqlite3; pipeline and setup operations also use jq.

The repository doubles as a Claude plugin marketplace. .claude-plugin/marketplace.json points to plugins/knowledge-distillery/, whose skills/ directory is also discoverable by portable skill installers.

Status

Proof of concept. The design is intentionally kept open to change — tight constraints and exhaustive safeguards make future iteration harder. Focus is on real-world usability first.

License

TBD

About

A system that delivers only verified knowledge to AI coding agents. 3-layer architecture with structural air gap.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages