wiki

wiki |

MCP server that maintains a persistent Obsidian-vault knowledge base.

Instead of answering from memory or re-reading the same files, you ingest once and query forever. Instead of searching across flat notes, you traverse a typed knowledge graph — thoughts, entities, reasons, questions, conclusions — all auto-classified by topic via OpenAI embeddings.

💡 Why

Every time an AI answers a question, it starts from scratch — re-reading files, re-summarizing context, re-deriving the same conclusions. This wastes tokens and loses compounding knowledge.

wiki fixes this by maintaining a single vault per project. Ingest a document once; the knowledge is chunked, embedded, purpose-classified, and indexed. Future queries hit the vault instead of raw files. Entity linking deduplicates overlapping content automatically.

Source = truth. .wiki/ = derived knowledge. Blow it away anytime; rebuild from source via /reindex.

⚡ Token savings

Operation	Without wiki	With wiki
Answer a recurring question	Re-read files every time	Single search_fulltext call
Cross-topic synthesis	Manual grep + summarize	search_by_tag + reason traversal
Code file read	Full file in context	Structure map + targeted body load

⚙️ How it works

source doc  →  ingest_thought / ingest_entity
                  ↓ OpenAI embed → purpose classification
            →  .wiki/thoughts/<uuid>.md   (type: thought, purpose: <tag>)
            →  .wiki/entities/<uuid>.md   (type: entity, purpose: <tag>)
            →  .wiki/reasons/<uuid>.md    (PartOf / Supports / ...)

.wiki/
  purposes/   — topical buckets (each has an embedding)
  thoughts/   — atomic facts from sources
  entities/   — recurring concepts
  reasons/    — directed edges between nodes
  questions/  — open questions
  conclusions/— synthesized knowledge
  ingest_log/ — append-only JSONL audit trail (ingest.jsonl, rotates at 265 lines)
  .search/    — Tantivy full-text index

• Purpose classification — every doc is embedded and matched (cosine) against purpose embeddings. Top-1 above wiki_similarity_threshold wins; below threshold → general.
• Chunking — multi-topic content splits into a parent container + child chunk docs linked via PartOf reasons.
• Entity linking — learn_pass / link_doc rewrites bare entity mentions as [[wikilinks]] and folds near-duplicate paragraphs into the canonical entity.

🛠️ Tools

All multi-doc tools are batch-only — wrap every payload in {items: [...]}, even for one record.

Tool	What it does
ingest	📥 Batch-write docs — kind: thought | entity | question | reason | conclusion. Auto-links, body-start [[<qid>]] → Supports (synthesis-fed), explicit reason_kind:"Answers" for direct answers
search	🔍 Batch hybrid search — mode: smart (conclusions-first) | fts (BM25) | tag | qa | list. raise_on_miss opt-in
get	📖 Batch-fetch docs with split inbound/outbound reasons + edge walk
update	✏️ Batch-update content/tags. Re-embeds + re-links on body change
retag	🏷️ Bulk add/remove tags + bulk-purpose-move without touching content
delete_doc	🗑️ Batch-delete by id/ids; cascades edge cleanup
learn_pass	🔁 Sensemaker — link/dedupe → connect → raise/answer → promote conclusions. limit:0 = scan whole vault. Returns invariant_reason when no progress
list_open_questions	❓ Paginate unresolved questions, filter by purpose
mark_question	✅ Batch-resolve questions (deleted | buried)
purpose	🏷️ Manage purposes — action: list | create | delete | reembed
author	🧹 Vault maintenance — action: reindex | sanitize | migrate | migrate_lifecycle | feedback | retitle_questions
code	💻 Code index ops — action: index | search | read | refs | validate
docs	📚 Fetch tool / concept markdown docs (no arg → list)

❓ Question lifecycle

Two states only:

• open — questions/<purpose>/.... Default for any new question.
• graveyard — questions/graveyard/<purpose>/.... Junk or unanswerable. Excluded from list_open_questions, learn_pass, and conclusions-first search.

Anything else is hard-deleted. An answered question is one that was promoted: a conclusion doc now exists with an Answers edge to the (former) question, and the question file itself is gone. The conclusion body preserves the original question text as a preamble so context survives the delete.

Resolve open questions via mark_question:

Status	Effect
deleted	Hard delete + cascade-delete every reason touching it. Use for already-answered or junk questions.
buried	Move to questions/graveyard/, tag graveyard, rewrite inbound wikilinks. Reversible — move the file back into the open tree to resurrect.

To re-explore everything in the graveyard: delete questions/graveyard/.