Skip to content

spillwavesolutions/research-knowledge-capture

v0.2.9MIT

Research Knowledge Capture (RKC): ingest deep-research dumps and articles into an OKF ContentPack of areas, subjects, claims, evidence, and findings. Layer 0 of the research second brain.

GitHubOpen Website

Research Knowledge Capture (RKC)

Layer 0 ContentPack for a research second brain. OKF Markdown + YAML is the source of truth. Agent Brain is a disposable index — see research-graph.

Version: 0.2.9 — rg↔scan parity holds on a symlinked root; parse-once rkc_search/rkc_validate, fail-closed find_rg. See the PRD.

What it owns

Eight nouns: ResearchArea, Subject, ResearchTask, SourceDocument, ResearchQuestion, Claim, Evidence, Finding.

Registered rels: has_subject, related_to, has_task, ingested_from, asks, answers, produced, asserts, evidenced_by, contradicts, supersedes, same_as.

Article bridge (Article → draws_from → Finding) is owned by content-media, not this pack.

Install

claude plugin marketplace add SpillwaveSolutions/research-knowledge-capture
claude plugin install research-knowledge-capture@rkc-plugin-marketplace

Hosts: Claude Code, Grok Build, Codex, Cursor, Agent Plugins 1.0. See docs/HOSTS.md.

Commands

HostIngestExtractPackRetrieveSearchValidateSpine
Claude / Grok / Cursor/research-ingest/research-extract/research-pack/research-retrieve/research-search/research-validate/research-spine
Codex$research-ingest$research-extract$research-pack$research-retrieve$research-search$research-validate$research-spine
python3 scripts/rkc_validate.py --root sample-knowledge
python3 scripts/rkc_pack.py subject.loop-policy.01J8X000000000000000000001 --root sample-knowledge --summary
python3 scripts/rkc_search.py "false civic alerts" --root sample-knowledge --limit 5 --json
python3 tests/test_rkc.py
python3 tests/test_extract.py
python3 tests/test_plugin.py
python3 tests/test_bulk_fixes.py
python3 tests/test_spine.py
python3 tests/test_search.py

Public samples are Northstar / Lumenfield fiction. Live dumps stay in private trees.

Bulk ingest

  • Lookup is research/catalogs/ingest-index.json (not a full-tree scan). Rebuild with --rebuild-index.
  • Archive-only ingest still creates the Subject and writes has_task. Pass --area <slug> for a ResearchArea + has_subject.
  • index.md / README.md titles are parent/filename. --extractor-version defaults to the current extractor, so omitting --extract on a re-run is idempotent.
  • Repair a 0.2.0 tree with python3 scripts/rkc_spine.py --knowledge knowledge --link-tasks then --area-map areas.json.
  • Heuristic extract skips files over 200 KB. Use --force-large only when you mean it.
  • One process per Subject (or per knowledge tree). Do not pkill -f a multiprocessing ingest; kill by PID.
  • vendor is free text. Convention: grok gemini claude deepseek chatgpt article perplexity unknown.
  • --source-kind defaults to deep_research.

Retrieval ladder

rkc_search (rg → scan) → research-retriever / rkc_pack --summary → BM25/Chroma → Kuzu last (Layer 1). Q&A retrieval is a sub-agent: the parent gets a summary card only. research-graph may feed seeds later and is not required for L0.

Actor

grok-bot/research-knowledge-capture. Isolated session. Ready PR. No force-push. No write to main.