SEC Filing Search
Ask a question about a US public company. Two embedding models search the same 1,422 passages from 10-K filings of 24 companies, and you see their top results side by side: my fine-tuned model and the original bge-small it started from. Companies marked unseen were held out of training, so they are the fairest test. The first four examples are real test questions where the two models disagree. Both models run entirely in your browser.
How well does it work? Benchmark, method and limitations
355 questions from 6 companies never seen in training (Chevron, Johnson & Johnson, Meta, Microsoft, Oracle, Walmart). Each question is ranked against all 355 passages from those companies; exactly one passage is correct.
| Metric | Base bge-small | Fine-tuned | Change |
|---|
How it was built: 10-K filings from SEC EDGAR, split into passages from Items 1, 1A, 7 and 7A; an LLM wrote one analyst-style question per passage; the split is by company to avoid leakage; fine-tuned with MultipleNegativesRankingLoss (in-batch negatives) on a T4 GPU. In this page the models run as int8 ONNX through Transformers.js, passage vectors are precomputed, and only your question is encoded live. Limitations: the training questions are synthetic; only narrative 10-K text is covered; free-form questions have no ground truth, so judge the passages by reading them. Similarity scores are on different scales for the two models, so compare passages, not numbers.