TeXposit Uses Jev to Speed Up AI Assistant Workflows

Overview

TeXposit, a specialized AI writing assistant designed for technical documentation and LaTeX projects, has integrated a new lightweight model called Jev to streamline its workflow. The update addresses a common bottleneck in AI-assisted writing: the initial overhead required to prepare for complex tasks.

Previously, assistants spent significant time browsing skills, loading project files, and reading tool schemas before generating any substantive output. By offloading this preflight phase to Jev, TeXposit can now parallelize data gathering, resulting in perceptibly faster response times without compromising the quality of the final output.

Accelerating Initial Workflows

The core innovation lies in how TeXposit handles the first few turns of a session. Traditionally, this phase involves serial operations that create a "time floor," where users wait passively while the system gathers context. The new approach introduces parallelism by using Jev to answer a batch of narrow, yes-or-no questions about the user's request.

For example, when asked to identify complex concepts in a paper that need TikZ visualization, the assistant previously spent 59 seconds on setup—loading the TikZ skill, reading project comments, and fetching specific file ranges—before answering in just 14 seconds.

With Jev, the system evaluates 42 questions regarding skills, deferred tools, and candidate files simultaneously. It identified the necessary TikZ skill and key files, while also flagging a secondary skill and a PDF download tool. Although this introduced a minor token waste, the overall time savings are significant, particularly when working with slower, more capable main models.

"We have not measured the aggregate time savings or benchmarks, but the difference is perceptible, especially when working with slower main models."

Crucially, Jev acts only as a preparatory layer. It does not have authority to act on the project directly. The main model retains full decision-making power, and any edits still pass through the same validation and approval processes as before. This ensures that speed improvements do not come at the cost of output integrity.

Judging Rigour with Speed

In a recent update, TeXposit introduced Rigour Mode, which automatically checks final answers against customizable requirements. While this feature improved output quality, it relied on a heavy LLM-based judge that incurred significant time and token costs. The judge typically reported which criteria failed without providing detailed explanations, making it a prime candidate for replacement by a faster model.

Jev now powers this judging process by testing each active requirement in parallel against the conversation history and the proposed answer. If a check fails, feedback is sent back to the main model for revision before the user sees the result.

Comparative testing revealed stark efficiency gains. In one example involving five evaluation cases, the old LLM-based judge took approximately 28 seconds, whereas Jev completed the same task in just 2.4 seconds. Jev agreed with the previous judge on four cases but rejected a good answer in one instance due to a lack of disclosed limitations.

The development team acknowledges that occasional false positives may occur. However, they argue that Rigour Mode is useless if disabled, and users are unlikely to enable it if it significantly slows down their workflow. To mitigate the risk of strict judges blocking tasks indefinitely, the system implements a correction cap.

After three failed attempts, the task proceeds with an explicit warning to the user that the answer may not meet all rigorous standards. This prevents impossible requests from halting progress and helps users recognize when their question cannot be answered with available evidence.

Turn-Level Routing and Skill Discovery

Beyond initial setup and judging, Jev serves as a turn-level router within the assistant. TeXposit employs a three-tiered model setup, ranging from fast and general-purpose models to smarter, more capable ones. Jev analyzes each user input to determine which tier is appropriate for the next response.

If Jev is uncertain about the complexity of a turn, it escalates to a higher-tier model. In cases where routing fails, the system defaults to the smartest available model. This dynamic allocation ensures that simple, mechanical tasks—such as fixing a compiler error or running a named tool—are handled quickly, while complex inquiries receive the attention of a more capable model.

Additionally, Jev has replaced hand-written phrase lists for automated skill discovery. Previously, the system used basic NLP to recommend skills based on explicit keywords in the user's prompt. Now, Jev evaluates whether a request points to a specific tool or skill, even if the user does not use the exact descriptive terms. This semantic understanding allows for more accurate skill selection and reduces the need for manual configuration.

The routing feature is currently excluded for users who provide their own OpenRouter API keys. In these cases, the system respects the user's choice of model, avoiding unsolicited switches that could degrade the experience despite potential time savings.

As TeXposit continues to refine its integration with Jev, the focus remains on balancing speed, cost, and reliability. By delegating preparatory and evaluative tasks to a specialized lightweight model, the platform aims to deliver a smoother, more responsive experience for technical writers without sacrificing the rigor required for professional documentation.