temporiki/
  raw/                    # Immutable sources
  wiki/                   # LLM-maintained compiled knowledge
    index.md              # Catalog
    log.md                # Chronological operations
    entities/
    concepts/
    sources/
    synthesis/
    decisions/
    queries/
    meta/                 # Dashboards (webclips activity, stale pages, etc.)
    _templates/           # Canonical page templates
  .temporiki/AGENTS.md               # Agent schema (Codex/OpenCode/Cursor)
  .temporiki/CLAUDE.md               # Claude compatibility mirror
  .temporiki/mempalace.yaml
  .memplite/palace.sqlite3 # Lightweight memory DB (created on demand)

One-time Obsidian setup:

1. Open Obsidian and create/select a vault from the repo folder.
2. Install/enable Obsidian Terminal and Web Clipper.
3. (Optional) confirm Web Clipper target folder is raw/webclips/ (Temporiki auto-sets Obsidian attachment path there).

Daily use:

1. Open Obsidian vault.
2. Open a terminal inside Obsidian, launch your LLM CLI (claude, codex, gemini, etc.), and chat.
3. Clip content to raw/webclips/; the Temporiki Auto-Ingest plugin fires within ~2s and indexes the new source.
4. Processed clips are moved to raw/webclips/_archive/YYYY-MM/ automatically.
5. Files dropped into raw/ via Finder or CLI while Obsidian is closed get picked up on the next palace-search / palace-kg-query via a lazy staleness check — no missing context.

No manual uv run ... commands are required for normal use.

uv --project .temporiki run temporiki ingest
uv --project .temporiki run temporiki onboard
uv --project .temporiki run temporiki palace-mine
uv --project .temporiki run temporiki palace-search "auth decision"
uv --project .temporiki run temporiki palace-kg-query --as-of 2026-04-13
uv --project .temporiki run temporiki lint
uv --project .temporiki run temporiki lint --autofix
uv --project .temporiki run temporiki query "What decisions are active?" --answer "..."

• Hash-based delta ingest tracked in .manifest.json
• Query save-back to wiki/queries/ (compounding knowledge)
• High-confidence auto-save from palace-search to wiki/queries/
• Wiki linting for missing frontmatter, broken links, and orphans
• Lint checks for pages missing from wiki/index.md
• Lint contradiction watch for overlapping active decisions on the same topic
• Built-in mempalace-lite (SQLite FTS5 memory search + temporal decision query)
• Thin Chroma integration via chromadb-client (HTTP mode)
• Hybrid retrieval reranking with confidence + citation rationale
• Intelligent query routing (KG -> Hybrid -> SQLite fallback)
• Pydantic + YAML schema validation with lint autofix support
• Context Graph mode guidance in .temporiki/AGENTS.md (wiki/decisions/ + temporal precedence)
• Session-launch hook for local Chroma Docker autostart
• In-vault Obsidian plugin (.temporiki/obsidian-plugin/) for eager ingest on vault create/modify events
• Router-lazy ingest guard (.temporiki/temporiki_tools/stale.py) ensures palace-search and palace-kg-query run against a fresh manifest even when Obsidian never opened
• Web Clipper inbox at raw/webclips/ with automatic ingest/index loop
• Webclips activity dashboard at wiki/meta/webclips-activity.md (inbox + recent archives)
• Lightweight default install: no chromadb/kubernetes stack unless --extra mempalace is requested

Most retrieval-augmented-generation systems rely on similarity in high-dimensional embedding spaces to pull context for the LLM. This has a well-known weakness — the curse of dimensionality: as dimensions grow, distances between points become less discriminating, and "semantically similar" can drift from "actually relevant."

A Stanford RegLab study of commercial legal RAG tools (Magesh et al. 2025, Journal of Empirical Legal Studies) measured hallucination rates of 17–33% on systems marketed as "hallucination-free." The paper identifies three failure modes that trace back to retrieval limitations:

1. Text-similar but irrelevant retrieval — shared surface tokens (a name, a word) pulling unrelated sources
2. Temporal drift — retrieved sources that were once correct but have been superseded
3. Misgrounding — citations to real sources that don't actually support the claim

Temporiki mitigates each of these with structural, not purely statistical, choices:

Temporiki does not eliminate hallucinations — no RAG system does. It narrows the surface area where the common failure modes live.

• SemVer tags are used (vMAJOR.MINOR.PATCH).
• .temporiki/pyproject.toml is the authoritative package version.
• .temporiki/CHANGELOG.md tracks release notes.
• GitHub Releases are auto-created when a SemVer tag is pushed.
• Version Guard CI fails if impactful runtime changes are made without a version bump and changelog update.

Release command:

./.temporiki/scripts/release.sh 0.1.3
git push origin main --follow-tags

For day-to-day work:

# Bump version based on change scope
./.temporiki/scripts/bump_version.sh patch   # bugfix/small behavior change
./.temporiki/scripts/bump_version.sh minor   # backward-compatible feature
./.temporiki/scripts/bump_version.sh major   # breaking change