Archeion

The name Archeion (say ar-KAY-on) is the Greek arkheion: the public records office of a Greek city, where official records and evidence were kept.

Archeion is a Tekmora project. Tekmora, from Greek tekmor (Homeric tekmor): a set goal, and the sign that shows the way to it; said TEK-mo-ra.

Extracted Cordana component providing the Knowledge Subsystem: the knowledge registry, MCP bridges, hybrid retrieval engine, snapshot compiler, evidence verification harness, external adapters, and cluster MCP serving endpoints.

Overview

• tools/knowledge-registry: Python & Go registry toolset for discovering, validating, snapshotting, and materializing content-addressed knowledge records, source acquisition contracts, channel policies, and independent review bundles. Enforces strict JSON schemas and cryptographic signers.
• tools/knowledge-mcp: High-performance Go Model Context Protocol (MCP) server providing knowledge retrieval, registry inspection, official channel discovery, and dependency diffing for AI agents, plus an in-process Codex bridge and MCP 0.1 acceptance harness.
• tools/knowledge-retrieval: Go hybrid retrieval engine implementing BM25F token ranking, reciprocal rank fusion (RRF), multidomain evaluation suites, gold span lane ranks, and blind query scoring.
• tools/knowledge-snapshot: Go snapshot compiler, change-pack generator, inventory indexer, and standing admission verifier for knowledge repositories.
• tools/knowledge-evidence: Go evidence compiler, verification harness, and tamper-evident reviewer signature validator.
• tools/knowledge-adapter-v2: Go adapter translating between external source formats, OCI registries, Git trees, and Cordana canonical knowledge records.
• tools/knowledge-dogfood-evidence: Python verification harness enforcing fail-closed assertions, admission gates, and cryptographic audit proofs for dogfood cluster environments.
• apps/ai/kb-mcp: FastMCP Python server and Helm chart exposing internal cluster knowledge and read-only research tools over MCP (SSE and Streamable HTTP transports), backed by Open WebUI and MCPO.

Provenance & Monorepo Extraction

• Source: a private cluster monorepo of the same owner
• Monorepo Source Paths:
  • tools/knowledge-registry
  • tools/knowledge-mcp
  • tools/knowledge-retrieval
  • tools/knowledge-snapshot
  • tools/knowledge-evidence
  • tools/knowledge-adapter-v2
  • tools/knowledge-dogfood-evidence
  • apps/ai/kb-mcp
• Monorepo Cut Commit: 1b8c489d46868ccf8103a144c250e5534f8d9ecf (origin/main on 2026-09-05)
• This Repository: https://github.com/cordanaLLM/knowledge

Directory paths are kept as they were in the monorepo. The history before the first commit here stays with the source and is not part of this repository.

Repository Structure

.
├── .gitea/
│   └── workflows/
│       └── ci.yml                     # Gitea Actions CI workflow
├── .gitignore
├── README.md                          # Root documentation (this file)
├── apps/
│   └── ai/
│       └── kb-mcp/                    # FastMCP server & Helm chart for cluster MCP
└── tools/
    ├── knowledge-adapter-v2/          # Format adapter for canonical records (Go)
    ├── knowledge-dogfood-evidence/    # Fail-closed admission & dogfood evidence (Python)
    ├── knowledge-evidence/            # Tamper-evident reviewer signature validator (Go)
    ├── knowledge-mcp/                 # Core knowledge MCP server & Codex bridge (Go)
    ├── knowledge-registry/            # Content-addressed registry & acquisition runner (Python/Go)
    ├── knowledge-retrieval/           # Hybrid BM25F + RRF retrieval engine (Go)
    └── knowledge-snapshot/            # Snapshot compiler & change-pack engine (Go)

Running Tests Locally

The extracted component contains Go test suites, Python test suites, and Helm chart linters.

1. Go Tests

Run unit tests directly in each module:

# Knowledge Adapter v2
(cd tools/knowledge-adapter-v2 && go test -v ./...)

# Knowledge Evidence
(cd tools/knowledge-evidence && go test -v ./...)

# Knowledge MCP (server, bridge, and acceptance tests)
(cd tools/knowledge-mcp && go test -v ./...)

# Knowledge Registry Go Runners
(cd tools/knowledge-registry/source-acquisition-plan-enumerator && go test -v ./...)
(cd tools/knowledge-registry/source-acquisition-runner && go test -v ./...)
(cd tools/knowledge-registry/source-acquisition-runner-v3 && go test -v ./...)

# Knowledge Snapshot
(cd tools/knowledge-snapshot && go test -v ./...)

# Knowledge Retrieval
(cd tools/knowledge-retrieval && go test -v -run 'Test(BM25|Dense|Current|Hybrid|Lifecycle|Paired|Situating|FineSpan)' ./...)

Note on conformance and integration tests: Certain tests in tools/knowledge-registry/p1-src03-acquisition-runner and tools/knowledge-retrieval depend on shared monorepo schemas (knowledge/registry/schemas/) and test vectors (testdata/canonical-json/v1/vectors.json). Neither is part of this repository yet, so those tests fail here.

2. Python Tests

Run decoupled Python test suites using pytest:

# Dogfood evidence fail-closed assertions (52 tests)
python3 -m pytest -q tools/knowledge-dogfood-evidence/tests/test_fail_closed.py

# KB-MCP proxy tests (11 tests)
python3 -m pytest -q apps/ai/kb-mcp/files/test_mcpo_proxy.py

Note on registry and server tools tests: tools/knowledge-registry/tests/ tests assert against monorepo schema directories (knowledge/registry/schemas/) and require rpds-py==2026.6.3. apps/ai/kb-mcp/files/test_server_tools.py asserts against container image dependencies (fastmcp==3.4.2, httpx==0.28.1, mcp==1.28.1, starlette==1.3.1).

3. Helm Chart Linting & Template Rendering

Lint and render the kb-mcp Helm chart:

helm lint apps/ai/kb-mcp
helm template kb-mcp apps/ai/kb-mcp > /dev/null

4. Public Tree Check

No tracked file may hold a home path, a private network address, a host name under a domain this repository used before it became public, or the product word the tool carried before it was named Archeion:

python3 -m unittest discover -s scripts -p 'test_*.py'
python3 scripts/check_public_tree.py

The first command runs the planted cases, the second reads every tracked file. Example addresses that upstream fixture files hold are listed, one per file, in scripts/public-tree-addresses.txt; an entry that matches nothing fails the check. The Public tree check job (.github/workflows/public-tree.yml) runs both on every pull request.

Licence

Copyright 2026 Lusoris.

The repository is under the European Union Public Licence, version 1.2 (EUPL-1.2; text in LICENSE). No release has been tagged yet, so the first release and every later one are under these terms.

Part	Licence
Code, configuration and test data written for this repository	EUPL-1.2
Documentation text: this README and docs/	CC-BY-SA-4.0
Upstream files and excerpts in the scorer-blind evaluation fixtures under tools/knowledge-retrieval/testdata/	the licence of their source: BSD-3-Clause (Go), Apache-2.0 (Kubernetes, Cilium)

REUSE.toml names the licence and the copyright holder of every file by path, and LICENSES/ holds each licence text. The REUSE lint job (.github/workflows/licence.yml) checks both on every pull request; reuse lint runs the same check locally.