wikilens 0.7.0

Agentic intelligence layer for Markdown/Obsidian vaults — RAG + evaluated metacognitive agents.

wikilens

An agentic intelligence layer for Markdown / Obsidian vaults. RAG + evaluated metacognitive agents, built in public.

Status: Alpha · P7 shipped (v0.7.0). ingest, query, audit, contradict, gap, and answer all work end-to-end on local Markdown vaults. See benchmark →

---

What it is

wikilens turns a folder of Markdown notes (an Obsidian vault, a Zettelkasten, any personal knowledge base) into a queryable, auditable, self-aware knowledge system.

It is:

• Local-first. Runs on your machine. Your notes never leave unless you explicitly call a remote LLM.
• Agent-based. Individual capabilities (link auditing, contradiction detection, gap finding) are isolated agents with measured performance.
• Evaluated, not vibes. Every agent ships with a labeled test fixture and a reported score. No "it seems to work."
• Markdown-native. Understands Obsidian-flavored syntax: [[wikilinks]], YAML frontmatter, callouts, embeds.

It is not:

• A hosted SaaS (v1 is local only)
• A note editor (bring your own — Obsidian, VS Code, plain text)
• A chatbot wrapper ("chat with your notes" is table stakes; this is the layer above that)

Why it exists

Second-brain tools have exploded, but most are storage-shaped, not thinking-shaped. They help you save notes, not reason across them.

The interesting question isn't "can an LLM answer a question about my vault" — that's solved. The interesting questions are:

• Where does my vault contradict itself?
• What should be there but isn't (gaps)?
• Which links are missing or wrong?
• What am I circling without naming (emergent concepts)?

Each of those is an agent. Each has a measurable success criterion. That's the project.

Who it's for

Knowledge workers who keep a serious Markdown vault (≥ 200 notes) and want more than search. Researchers, writers, developers, second-brain practitioners.

Roadmap

Shipped: P1 – P6 (RAG core, Link Auditor, Contradiction Finder, Gap Generator, Answer Generator, all with hand-labeled evals).

Next: P7 — PyPI + installer polish, then P8 — Temporal Drift Detector, P9 — Unnamed Concept Detector, P10 — Epistemic Confidence Mapper, P11 — Obsidian plugin, P12 — v1.0 launch.

Full phase list, launch hooks, and eval targets in ROADMAP.md.

Design principles

1. No silent steps. Every agent explains what it did and why.
2. Reproducible evaluation. make benchmark produces the numbers in BENCHMARK.md.
3. No vendor lock. Swappable embeddings, swappable LLMs, swappable vector stores.
4. Fail loud. Broken inputs are surfaced, never guessed.

Install

pip install wikilens

Python 3.12+ is required. The first run of ingest or query downloads two local models (~270 MB total, cached for all subsequent runs): BAAI/bge-small-en-v1.5 (embedder) and BAAI/bge-reranker-base (reranker).

For contradict, gap, and answer you also need a remote LLM key:

pip install 'wikilens[judge]'         # adds openai, anthropic, scikit-learn
export OPENAI_API_KEY=sk-...          # or ANTHROPIC_API_KEY for --judge claude

From source (dev / contributor install):

git clone https://github.com/Universe8888/wikilens.git
cd wikilens
pip install -e '.[dev]'

On Windows, if the wikilens command is not found after install, add the Python Scripts directory to PATH (e.g. %APPDATA%\Python\Python312\Scripts) or run python -m wikilens.cli while developing.

Usage

# Build the index (full rebuild each run; incremental is deferred).
wikilens ingest ./my-vault

# Query — four retrieval modes are supported.
wikilens query "how do plants turn light into sugar"            # default: rerank
wikilens query "..." --mode dense                               # cosine only
wikilens query "..." --mode bm25                                # FTS / BM25 only
wikilens query "..." --mode hybrid                              # RRF fusion
wikilens query "..." --mode rerank -k 10                        # top-k after rerank

# Audit — find broken, one-way, orphan, and shadowed wikilinks.
wikilens audit ./my-vault                                       # markdown report
wikilens audit ./my-vault --json                                # machine-readable
wikilens audit ./my-vault --only broken,orphan                  # filter classes

audit exits 0 when clean, 1 when any finding is reported — so it doubles as a pre-commit / CI gate. Index defaults to .wikilens/db inside the current directory; override with --db <path>.

# Contradict — find conflicting chunk pairs.
pip install -e '.[judge]'
wikilens contradict ./my-vault --judge claude                    # full run
wikilens contradict ./my-vault --judge none                      # dry-run (no API)
wikilens contradict ./my-vault --judge claude --sample 20        # cap API calls

# Gap — find unanswered questions the vault implies but doesn't answer.
wikilens gap ./my-vault --judge claude                           # full run
wikilens gap ./my-vault --judge none                             # dry-run (no API)
wikilens gap ./my-vault --judge claude --max-clusters 10         # budget cap
wikilens gap ./my-vault --judge claude --top-gaps-per-cluster 2  # fewer per cluster

# Answer — for each gap, retrieve vault evidence and draft a note stub.
wikilens gap ./my-vault --judge openai --json > gaps.json        # generate gaps first
wikilens answer ./my-vault --gaps gaps.json --judge openai       # draft stubs to stdout
wikilens answer ./my-vault --gaps gaps.json --judge openai \
    --write --out ./stubs/                                       # write .md files
wikilens answer ./my-vault --gaps gaps.json --judge none         # dry-run (no API)

contradict, gap, and answer exit 0 when clean, 1 when findings / partial coverage reported. answer exits 2 on bad input or file collisions when --write is set. Set OPENAI_API_KEY (or ANTHROPIC_API_KEY for --judge claude) in your shell or in a .env file at the repo root.

Benchmark

Four eval suites, all reproducible from a fresh clone.

Retrieval (P2): Hit@5 = 1.00 in all four modes on the 36-note synthetic vault. Latency p95: 37 ms (dense) to 1846 ms (rerank).

Link audit (P3): Precision = Recall = F1 = 1.00 on all four detector classes (16-note fixture, 19 planted defects).

Contradiction finder (P4): F1 = 0.82, retrieval recall = 0.90 on the 24-pair hand-labeled fixture. Wall clock 67.7s.

Gap generator (P5): Cluster-stage recall = 1.00, matcher-stage F1 = 0.65 on 10 gold gaps. All 10 gold gaps surfaced by some cluster.

Answer generator (P6): Pass rate = 0.80 (8/10 drafts pass all 4 axes: faithfulness, coverage, attribution quality, stub structure). Attribution rate = 1.00 (automated). Wall clock 90s for 10 gaps.

See BENCHMARK.md for full tables. Reproduce any suite:

# P6 answer generator
pip install -e '.[dev,judge]'
rm -rf .wikilens_p5_eval
wikilens ingest fixtures/gaps_vault --db .wikilens_p5_eval/db
python scripts/eval_p6.py --judge openai

Local checks:

make lint
make typecheck
make test
make benchmark

make benchmark uses no-API mock judges for P4-P6 by default. To reproduce published LLM-judged numbers, run the individual eval scripts with --judge claude or --judge openai after setting the relevant API key.

Writing / research

Design decisions and methodology writeups are published as they happen. Index will live at /docs once P2 is done.

License

MIT — see LICENSE.

Author

Built by Boris Manzov. Feedback and ideas welcome in Issues.