01 · A
Problem
Search can retrieve a note without helping a learner practise or retain it. The product needed to connect answers to evidence and learning actions without silently turning model output into truth.
All projects07 / 09
Local-first active learning
Vault Coach
An Obsidian learning companion that connects source-grounded retrieval, practice, mastery feedback and a reviewable concept-evidence graph.
- Period
- 2026 — present
- Focus
- Knowledge Systems · Agentic AI · Software Engineering
- Core stack
- TypeScript · Obsidian API · RAG · Knowledge Graphs
TL;DR
- Problem
- Search can retrieve a note without helping a learner practise or retain it. The product needed to connect answers to evidence and learning actions without silently turning model output into truth.
- What I did
- Implemented source-linked Q&A, assessment workflows, mastery views and review recommendations.
- Result / limit
- As of October 2026, the Obsidian plugin marketplace reported about 1,000 downloads and 158 regular users; GitHub reported 9 stars.
01 · B
Context
The plugin runs inside Obsidian, adapts to vault files and local stores, and uses a bundled multilingual embedding model in an isolated WASM worker.
Why it matters
A learning agent is useful only when learners can trace an answer to evidence, practise recall and decide which machine-proposed relations become trusted knowledge.
Explore technical detailsWhat made it hard · System path · Key trade-offs
What made it hard
- 01
Keep user knowledge in the vault unless a configured model call needs selected context.
- 02
Preserve source locations across Markdown and text-based PDF chunks.
- 03
Review model-proposed semantic relations before they affect the effective graph.
- 04
Recover interrupted local writes and invalidate stale derived data.
My approach
System path
Index
Parse vault content and maintain incremental text/vector indexes
Retrieve
Fuse keyword and vector rankings while preserving source locations
Learn
Generate exams, record assessment facts and derive mastery views
Govern
Confirm, reject or undo proposed concept-evidence relations
Grounded retrieval
Two rankings meet only after each has done its own job
Keyword search captures exact language, vector search captures semantic similarity, and reciprocal-rank fusion combines rank positions without pretending the raw scores share one scale.
01query
Question
User intent and current context
02lexical
Keyword rank
Exact terms and document language
03semantic
Vector rank
Bundled local embedding representation
04hybrid
RRF
Fuse rank positions, not raw magnitudes
05traceable
Answer + sources
Return passages with page/note links
Key trade-offs
RRF hybrid retrieval
Keyword and vector rankings stay independent before reciprocal-rank fusion; raw scores are not treated as one scale.
Candidates are not facts
Deterministic structure, model proposals, user decisions and the effective graph remain separate layers.
Facts before projections
Assessment records are authoritative; mastery and recommendations are reproducible derived views.
04 · A
My contribution
- Implemented source-linked Q&A, assessment workflows, mastery views and review recommendations.
- Built the concept-evidence graph with confirm, reject and undo gates for model-proposed relations.
- Added incremental indexing, local integrity checks, backup recovery, tests and CI assets.
04 · B
What was implemented
- Implemented an end-to-end learning loop from evidence retrieval to practice and governed graph updates.
- Kept AI suggestions separate from approved graph facts through explicit user decisions.
- As of October 2026, the Obsidian plugin marketplace reported about 1,000 downloads and 158 regular users; GitHub reported 9 stars.
Graph governance
Relationship extraction is a proposal workflow, not an automatic fact write
The effective graph contains deterministic structure and explicitly accepted relations. Rejected and pending proposals remain inspectable without becoming trusted knowledge.
01source facts
Deterministic structure
Documents, headings and explicit links
02pending
Model proposal
Candidate concept or evidence relation
03governance
User decision
Confirm, reject or undo
04trusted
Effective graph
Only accepted relations affect learning views
Lessons learned
01Local-first describes data ownership, not a promise that every configured model stays offline.
02A knowledge graph becomes trustworthy through governance states, not by drawing more edges.
What I can share
This is my open-source Obsidian project. Its production path uses WASM rather than WebGPU. The public marketplace and GitHub figures describe reach, not learning effect.
Public sources