Build and personalize small software that runs your work, on top of your company brain.
Teams use Paprwork to fork an app someone already built, make it theirs, and let agents run it on a schedule. Every app reads from and writes to one memory graph that you own and that works with any model.
papr.ai · Platform and developer docs · GitHub · Evals
The company brain is also a platform you can build on. The Papr SDK and API give your own agents and apps memory and context intelligence: schemas, policies and retrieval, portable across models. Get an API key and read the developer docs at platform.papr.ai.
We train our own embedding models for that layer. There are two series, and you use them in different ways:
| Series | What it is | How to use it |
|---|---|---|
| papr-embed-code-1 (4B, 0.6B) | Retrieval for code and text. Apache 2.0, fp32 reference precision. | Download the open weights here, or call it through the Papr API. |
| papr-embed-schema-1 (4B, 0.6B) | Schema-conditioned embeddings. You describe what matters in your domain as a schema, and the vector carries it as explicit, inspectable dimensions on top of a base embedding. | Use it through the Papr API and SDK at platform.papr.ai. The weights are not released. |
Pick code-1 if you want weights you can run yourself. Pick schema-1 if you want to control what "similar" means for your data, by changing the schema instead of retraining.
| Model | Params | CoIR (10 tasks, nDCG@10) |
|---|---|---|
| papr-embed-code-1-4b | 4B | 80.13 |
| papr-embed-code-1-0.6b | 0.6B | 77.19 |
Stock mteb 2.10, fp32, max length 8192. We trained on the train splits of most CoIR tasks, so this is mostly an in-domain result. Each model card lists what was and wasn't seen, the batch size per task, and the full evaluation protocol. These are code models and are flat to worse than the base on non-code retrieval.
| Model | SciFact (nDCG@10) | vs. its base |
|---|---|---|
| papr-embed-schema-1-4b | 78.78 | +0.44 over Qwen3-Embedding-4B |
| papr-embed-schema-1-0.6b (SciFact-pinned) | 72.99 | +3.27 over Qwen3-Embedding-0.6B |
Official MTEB SciFact test split, served through the Papr API.
| Repo | What it is |
|---|---|
| Qwen3-Embedding-4B-CoreML | Qwen3-Embedding-4B converted to Core ML for on-device use on Apple hardware. |
Start with the developer docs, or see the product at papr.ai. For questions, open an issue on GitHub.