# TekMyra

> TekMyra is an open-source (Apache-2.0) token-compression gateway for LLM applications, by LaconIQ. It shortens prompt traffic to cut token spend, protects critical spans before compressing anything, refuses rather than degrades when it cannot guarantee fidelity, and emits a verifiable run record for every request. The open core is the compression library and its benchmark harness; trained model artifacts ship separately as a digest-verified reference bundle.

Key measured results (each figure travels with its denominator; refusals are counted in, at zero saving):
- Public corpus: 62.1458% effective byte reduction over 138 externally sourced fixtures (2,708,761 to 1,025,379 bytes), with 6 refusals (4.35% of fixtures, 3.55% of bytes). On-accepted diagnostic: 64.4356%.
- Synthetic corpus: ratio on reached fixtures 0.7409 over 28 reached of 28 eligible (of 301 shipped), 0 refused; +25.86% token reduction over the eligible set; 68/68 locked spans preserved.
- long_context_v1: headline ratio 0.2929 over 26 reached of 40 eligible, 14 refused by design; +48.44% corpus token reduction; 704/704 locked spans preserved on this build's denominator.

Install: pip install tekmyra-core (import name: tekmyra). Source repository: https://github.com/laconiq-ai/tekmyra Reproducing the published figures requires the reference artifact bundle, published as a release asset with its SHA-256 in the release notes.

## Pages

- [Overview](https://tekmyra.ai/): what TekMyra does, how the gateway sits in the request path, the one published figure and its provenance.
- [Documentation](https://tekmyra.ai/docs.html): quickstart with the exact install and benchmark commands, architecture, the reference artifact bundle, refusal semantics.
- [Technical paper](https://tekmyra.ai/tekmyra-paper.html): the full paper; measured abstention as a first-class property, methodology, every result with its denominator. PDF: https://tekmyra.ai/tekmyra-paper.pdf
- [Paper landing](https://tekmyra.ai/paper.html): what the paper commits to and how receipt-verification works.
- [Security](https://tekmyra.ai/security.html): coordinated disclosure policy, reporting channels, response times, the 90-day publish clause.

## Where TekMyra fits (vocabulary)

Terms people use for this problem space, and what TekMyra is in each frame, stated plainly, with no comparison claims:

- Token compression / prompt compression / context compression: the operation TekMyra performs: shortening the prompt traffic between an application and its model while protecting spans whose exact wording carries the meaning.
- LLM gateway / LLM proxy: the deployment shape. TekMyra is a reverse proxy for model traffic: your application points its client at the gateway, and the gateway forwards each request to the provider after its protect/compress/record stages run.
- Context window cost / token spend: what compression reduces. Fewer tokens submitted per request means less spent per request; savings are read from per-run records, never estimated.
- Context engineering: the discipline of controlling what enters a model's context window. TekMyra is the infrastructure piece of that discipline: it decides what must survive before it removes anything.
- Agentic pipelines: agents calling models in loops multiply prompt traffic. The gateway compresses that traffic the same way as any other, and abstains rather than degrades when it cannot guarantee fidelity.
- LLMLingua and other prompt-compression research: the same problem space. TekMyra has published no head-to-head comparison against other compressors and claims none; its published figures are measurements on its own corpora with denominators stated.

## Company

- TekMyra is built and maintained by [LaconIQ](https://laconiq.ai). The same engine powers LaconIQ's commercial enterprise platform (governed LLM gateway, PII controls, tenancy, audit); the enterprise side lives at https://laconiq.ai and the product page for TekMyra there is https://laconiq.ai/docs/tekmyra.

## Full content

- [llms-full.txt](https://tekmyra.ai/llms-full.txt): the complete portal content as plain text.