Nuggetindex

NuggetIndex: time-aware, conflict-aware fact governance

You apply the NuggetIndex method (Zerhoudi et al., SIGIR '26): represent knowledge as atomic facts ("nuggets") with explicit temporal validity, detect when sources disagree, and surface conflicts instead of silently picking a winner.

Core model

A nugget is one atomic fact:

• Triple: (subject, predicate, object) + the original supporting text span
• Validity: interval [start, end) — end = null means "still true as far as we know"
• Lifecycle: active (current) | deprecated (obsolete) | contested (sources disagree)
• Provenance: which source(s) assert it

Three failure modes this prevents:

1. Temporal staleness — answering "Who is Apple's CEO?" with Steve Jobs because an old

passage ranked highest. Every fact carries [start, end); check it against query time.

1. Silent conflicts — two sources claim different values for the same single-valued

attribute; naive RAG picks whichever embeds better. Mark both contested and tell the user.

1. Rename drift — Twitter→X, Facebook→Meta. Track renames as facts (renamedTo) so both

surface forms resolve to one entity.

Decision ladder — cheapest rung that fits

Rung 1: In-context audit (most common — no files, no install)

The user shows you retrieved passages, an answer that looks outdated, or asks "is my RAG returning stale facts?":

1. Establish the query time (default: today; the user may ask about a past time).
2. Extract the fact triples relevant to the query from each passage, with any temporal

cues — follow [references/extraction.md](references/extraction.md).

1. Infer each fact's validity interval from its temporal cues (rules in extraction.md).
2. Group facts by key (subject, predicate). Apply the conflict rules in

[references/conflict-rules.md](references/conflict-rules.md): single-valued predicate + overlapping validity + different objects ⇒ contested.

1. Report per passage: consistent / stale (validity ended before query time) /

contested (disagrees with another passage), with the evidence spans.

Never resolve a contest yourself when evidence is symmetric — show both claims, their sources, and their intervals, and let the user adjudicate.

Rung 2: Project fact ledger (ongoing work, small working set)

The user wants governed facts across a session or project (docs, research notes, a small knowledge base): maintain nuggets.jsonl — one nugget per line in the exact schema of [references/ledger-format.md](references/ledger-format.md). The format is identical to the pip package's Nugget JSON, so the ledger imports into a real index later with no migration.

On every append: extract → canonicalize entity names (resolve known renames/aliases to one canonical form) → infer validity → dedup (same subject|predicate|object|start|scope = same fact; merge provenance) → run conflict detection against existing ledger entries and update lifecycle states. Report any new contested keys to the user immediately.

Rung 3: Escalate to the pip package (scale, persistence, serving)

Switch when any of these hold: corpus beyond ~50 documents, need for persistent hybrid BM25+dense retrieval, an HTTP sidecar for a production RAG stack, LangChain/LlamaIndex/ Haystack integration, or reproducible evaluation.

pip install nuggetindex        # extras: [openai] [anthropic] [dense] [serve] [all]
nuggetindex build ./corpus --db index.db
nuggetindex query "CEO Apple" --db index.db --time 2010-06-01

Full CLI/API surface, extras, and a hand-ledger import recipe: [references/package.md](references/package.md). LLM-backed extraction needs an API key (OPENAI_API_KEY / ANTHROPIC_API_KEY / GOOGLE_API_KEY, or local Ollama); the rule-based extractor works with none. If no key is available and rule-based extraction is too weak for the corpus, say so — do not fabricate extraction results.

Rules that always apply

• Timezone-aware UTC datetimes everywhere. Intervals are half-open [start, end).
• Query-time semantics: a fact answers a query at time t iff its validity contains t

and its status is active or contested — never deprecated.

• Surface, don't hide: contested facts are presented as disputes with provenance, e.g.

Apple CEO: Tim Cook (2014–present) [DISPUTED: source B claims John Sculley (overlapping)].

• Be literal in extraction: no invented facts, no paraphrased evidence spans, omit facts

you are not confident about.

• Unknown predicates default to multi-valued (conservative — avoids false conflicts).

Only the single-valued predicates listed in conflict-rules.md trigger contests by default.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: searchsim-org
• Source: searchsim-org/nuggetindex-skill
• License: MIT
• Homepage: https://nuggetindex.searchsim.org/

Install and usage instructions live in the source repository linked above.