kannaka-encoder
The text encoder that kannaka-memory loads when it embeds natively, with no Ollama or other service running.
What this is, plainly
Revision v0 is the stock sentence-transformers/all-MiniLM-L6-v2 weights, converted to fp16 safetensors. Nothing was trained or changed. It exists so the memory system has one pinned, hash-checked file to download, and so later revisions with our own edits have a baseline to be measured against. Credit for the model belongs to its original authors.
| Architecture | BERT, 6 layers, 384 hidden, 12 heads (22.7 M parameters) |
| Output | 384-dim, mean pooling over tokens, L2-normalised |
| Window | first 255 tokens plus [SEP] |
| Activation | GELU, tanh approximation |
| File | model.safetensors, fp16, 45 MB |
The pooling, window and activation above are chosen to match what Ollama's all-minilm does, so a store written through Ollama can be read with this encoder.
Measured agreement with Ollama all-minilm
200 varied texts, cosine between this encoder (run natively in Rust) and Ollama 0.34.4 for the same text. Measured 2026-11-14 on one Windows laptop.
| Pair | Min cosine | Median cosine | Nearest neighbour identical |
|---|---|---|---|
| Ollama CPU vs Ollama GPU (Ollama's own noise) | 0.999936 | 0.999986 | 198 / 198 |
| This encoder vs Ollama GPU | 0.999937 | 0.999986 | 198 / 198 |
| This encoder vs Ollama CPU | 0.9999986 | 0.9999995 | 198 / 198 |
Known gaps.
- Two of the 200 texts are excluded above and do not agree: a Korean text (0.659) and a Devanagari text (0.964). Ollama uses llama.cpp's WordPiece tokenizer, which treats those scripts differently from the Hugging Face tokenizer shipped here.
- Top-5 neighbour order matches in 188 to 194 of 198 cases; every difference is a swap between near-ties.
- An int8-quantised version was tried and rejected (median cosine 0.992).
- One machine, one Ollama version, 200 texts.
- Checked against one real store (2026-11-14). A 128 MB store written with Ollama
all-minilm, 100 English queries, top 10, query embedded by Ollama versus by this encoder (kannaka-memory v0.17.0): top result identical for 100 of 100; same ten memories for 97; same ten in the same order for 92 (Ollama against itself: 100, 100, 100). Every difference is a swap between near-ties (largest score gap on shared results 0.0018), and in seven of the eight differing lists the first difference is at rank 6 or lower. A Korean query shared only 3 of 10 results and a Devanagari query 8 of 10, both with a different top result (the tokenizer gap above). The queries were written for the test, not collected from use, and it is one store.
The probe and full tables are in the kannaka-memory repository under spikes/native-embed/.
Files
model.safetensors (fp16 weights), vocab.txt, tokenizer.json (stock; a loader must switch off its padding and its 128-token truncation), SHA256SUMS.
Roadmap
Later revisions will carry weight-space edits tuned for wave-interference recall. Each will have its own revision tag and its own measured comparison against v0, and none will be presented as compatible with v0 vectors.
Licence
Apache-2.0, the same as the base model.
Card written by Claude Code (an AI agent) for Kannaka Labs.
Model tree for Kannaka/kannaka-encoder
Base model
nreimers/MiniLM-L6-H384-uncased