Supported

Tags and links¶

Behavior¶

Link source	Strength	Use
Explicit document reference	Highest	Preserve authored relationships
Managed tag overlay	Policy-controlled	Add vocabulary and preserve source metadata
Similarity-derived link	Scored and bounded	Discovery. Review before treating as identity

On LongMemEval, the measured A-MEM threshold produced link precision 0.436 and recall 0.944: useful for discovery, too noisy for automatic equivalence.

Example tag overlay example

Source tags	Managed change	Effective tags
`policy`, `draft`	Add `reviewed`. Remove `draft`	`policy`, `reviewed`
`policy`	Add unknown tag	Rejected before mutation

Managed tags

TagDefinition, TagAssignments, assign_tags, normalize_tag, and search_weight store curation in an overlay. A provider resync preserves that overlay.

Derived links

extract_explicit_links finds explicit references. derive_links adds bounded similarity links and produces typed LinkCandidate values.

How it works¶

Tag keys are normalized before add/remove set operations. Definitions validate that assignments refer to known tags. Search weight combines assigned tag weights through deterministic policy and leaves source relevance intact. Explicit-link extraction recognizes source references first. Similarity linking scores caller-supplied candidate IDs, removes self-links, applies a threshold and limit, and sorts ties stably. Returned links are proposals. The application commits and interprets them.

curation.py¶

from mari_kit.knowledge import (
    TagAssignments, TagDefinition, assign_tags, derive_links, search_weight,
)

definitions = {"canonical": TagDefinition(key="canonical", label="Canonical",
    kind="canonical", search_weight=2.0, behaviors=("Wins conflicts",))}
assignments = assign_tags(TagAssignments(), doc.document_id, definitions,
    add=("canonical",))
weight = search_weight(doc.document_id, assignments, definitions)
links = derive_links(doc.document_id, candidate_ids,
    score=lambda source, target: similarity_scores[source, target])

Function definitions and options¶

Function	Required inputs	Caller-controlled options
`assign_tags`	Prior assignments, document ID, definitions	Explicit add/remove keys
`normalize_tag`	One raw key	Deterministic key normalization
`search_weight`	Document, assignments, definitions	Weights live in caller-created definitions
`extract_explicit_links`	Source document and known IDs	Recognized authored references
`derive_links`	Source ID, candidate IDs, score callback	Threshold and result limit. Self-links are removed

Research basis

Mari’s managed-tag overlay and bounded link proposal rules are deterministic curation contracts.

Record managed tags as an explicit input when they affect a derived ranking or artifact. The shared dependency planner can invalidate those outputs independently of source-text embeddings.