Contradiction Finder

Finds pairs of notes in your vault whose claims contradict each other, using embeddings to find related notes and an LLM to verify genuine conflicts. Knowledge bases drift: two notes written months apart can end up asserting opposite things ("we deploy on Fridays" vs. "we never deploy on Fridays"). This plugin surfaces those conflicts so you can reconcile them.

Features

• Four-stage scan pipeline:
  1. Represent — every markdown note is reduced to its title plus the first ~1500 characters of body text (configurable).
  2. Embed — notes are embedded in batches of 50 via the configured embedding model. Vectors are cached in plugin data keyed by file path + mtime, so re-scans only re-embed notes that changed.
  3. Pair — cosine similarity is computed across all note pairs; only pairs above the similarity threshold (default 0.80) are kept, since unrelated notes cannot meaningfully contradict. The most-similar pairs are capped (default 40) to bound LLM cost.
  4. Verify — each candidate pair is sent to the chat model, which returns a strict-JSON verdict: { contradiction, confidence, statementA, statementB, explanation }. Only confirmed contradictions are kept.
• Side panel — one card per confirmed contradiction: both note titles (click to open the file), the two conflicting statements, an explanation, and a confidence percentage. A Rescan button re-runs the pipeline.
• Persistent results — results survive restarts until the next scan.
• Works with any OpenAI-compatible endpoint — OpenAI, Ollama, LM Studio.

How to use

1. Open the plugin settings and enter your LLM base URL and API key (leave the key empty for local servers such as Ollama or LM Studio).
2. Run the Scan vault for contradictions command, or click the compare ribbon icon and press Rescan in the panel.
3. A progress notice reports each stage (embedding, pairing, verifying).
4. When the scan finishes, review the cards in the Contradictions panel; click a note title to jump to that note. The Open contradictions panel command reveals the panel at any time.

Settings

Limitations

• Only the first ~1500 characters of each note are considered; contradictions buried deeper in long notes may be missed.
• Similarity + verification are pairwise; a claim spread across three or more notes is not detected as a group.
• Verdicts are only as reliable as the chosen model; confidence is the model's self-report, not a calibrated probability.
• Large vaults produce O(n²) similarity comparisons in memory; the LLM cost is bounded by maxPairs, but embedding cost scales with the number of notes.
• Embeddings are cached per file (keyed by mtime and embedding model); changing the embedding model automatically re-embeds all notes on the next scan, which can be slow/costly on large vaults.
• Note text is sent to the configured LLM endpoint; do not scan a vault whose contents may not leave your machine unless the endpoint is local.

Privacy and network use

This plugin only makes network requests to the LLM endpoint you configure in its settings (OpenAI-compatible; a local Ollama or LM Studio works fully offline). Note titles and the first part of each note body are sent to that endpoint for embedding, and candidate note pairs are sent to the chat model for verification. Nothing is sent anywhere else, no telemetry is collected, and no requests are made until you run a scan with a configured endpoint. Using a hosted provider may require a paid account with that provider.

Installation

This plugin is not yet in the community store. To install it manually:

1. Build the monorepo so main.js is produced from src/main.ts.
2. Copy manifest.json, main.js, and styles.css into <your-vault>/.obsidian/plugins/contradiction-finder/.
3. Enable Contradiction Finder in Settings → Community plugins.
4. Open the plugin settings and enter your LLM base URL / API key.
5. Add a couple of notes that disagree (e.g. one saying "The office opens at 8am", another saying "The office opens at 9am"), then run Scan vault for contradictions and open the panel.

Support

If this plugin is useful to you, you can support its development at Buy Me a Coffee.