@muninhq/config-personal — the personal-knowledge configuration

The prosumer default configuration (local:init → ingest → extract → MCP): a conservative entity/relationship schema tuned for personal prose — meeting notes, journal entries, reading notes, project logs — with one all-access owner role and a single private sensitivity class.

Shape

| piece | choice | why | | --- | --- | --- | | Entity types | Person / Project / Topic / Source | precision over coverage (the mat-hr F12 lesson); four types cover what personal notes actually contain | | Relationships | worksOn (Person→Project) · about (Source|Project→Topic) · authoredBy (Source→Person) | fromTypes/toTypes make wrong edges unrepresentable | | Roles | one owner with baseTags: ['personal'] + both capabilities | a personal memory has exactly one human | | Sensitivity | one private class, tag personal, default | nothing to widen access to | | Tag expansion | flat identity | no hierarchy |

Alignment invariant: the owner role's base tags include personal — the exact tag local:init prints in its suggested ingest command — so the MCP server's union-of-baseTags read context sees everything that command writes. A cross-check test in munin-mcp (local-init-config-alignment.test.ts) asserts this, so the two can never drift.

Few-shot discipline

Few-shots are the quality lever. They are authored from fictional personal prose disjoint from the eval corpus, and each over-extraction-prone type carries an "extract nothing" example on realistic chatty prose — restraint is taught, not hoped for. Two personal-prose facts the examples encode:

1. People appear as first names only; the name as written is the identity.
2. The writer narrates in first person — "I", "my sister", "the plumber" are never entities.

Relationship examples live inside entity few-shots: the extraction prompt renders entity few-shots only, so an example on a relationship type would never reach the model.

Eval

src/eval/ ships a 9-document synthetic corpus (entirely invented content), a hand-authored ground-truth manifest, and a pure scorer. The live runs are providers-gated suites in munin-mcp (personal-eval.providers.test.ts): a cloud leg (Haiku) and a local leg (Ollama, qwen2.5:7b). Numbers and known limits: EVAL-FINDINGS.md.

Connector seam (future work — deliberately not here)

Markdown knowledge-base tools have quirks this configuration does not handle: [[wikilinks]], YAML frontmatter, embedded queries, tag syntax. Those belong to a future connector that normalises such files before ingestion (the connector tier — packages/connectors/*), not to this configuration package. This package assumes plain prose paragraphs, which is what the engine's ingestion pipeline produces today.

Open core

Part of the Munin open-core local product, released under AGPL-3.0-only (see LICENSE and the repository NOTICE). The hosted / managed product and the vertical configurations are a separate, closed commercial product and are not licensed under the AGPL.