Jev AI: the TypeSafe Jev model

Model routing: pick the cheapest model that can do the job

Let Jev read the request and choose a model by criteria you define: fast and inexpensive for lookups and extraction, more capable for architecture and high-stakes calls. LangChain ships an experimental ModelRouterMiddleware for this, and the probabilities stay available in agent state.

from langchain.agents import create_agent
from langchain_typesafe.experimental.middleware import (
    ModelChoice, ModelRouterMiddleware,
)

router = ModelRouterMiddleware(
    choices={
        "fast": ModelChoice(model="<cheap-model>",
                            criteria="Lookups, extraction, small edits."),
        "strong": ModelChoice(model="<strong-model>",
                              criteria="Architecture, high-stakes calls."),
    },
    instructions="Pick the cheapest model that can finish the task.",
)
agent = create_agent("<default-model>", middleware=[router])

LangChain guide to Jev

Agent guardrail: block risky tool calls before they run

Agents can be steered by bad instructions or by an attacker. Coding harnesses already classify dangerous actions before running them, and a cheap classifier lets you add the same check to any agent. LangChain's experimental AutoModeMiddleware uses Jev to inspect tool calls and block risky ones before the tool executes.

from langchain.agents import create_agent
from langchain_typesafe.experimental.middleware import AutoModeMiddleware

guardrail = AutoModeMiddleware(tools=["bash"])
agent = create_agent("<your-model>", middleware=[guardrail])

LangChain guide to Jev

Ticket triage and intent routing

One request can return the team (Choice), the frustration level (Score) and whether the message is urgent (Noul). Your code then decides: auto-reply, escalate or hand to a human. TypeSafe documents intent routing and confidence-gated routing patterns for exactly this.

res = client.system_one(
    state=ticket_text,
    questions={
        "team": Choice(instructions="Who should handle this",
                       criteria={"billing": "Payments", "technical": "Bugs"}),
        "urgent": Noul(instructions="Needs a reply today"),
    },
)
if res.answers["urgent"].noul > 0.9:
    escalate(ticket_text)

Intent routing pattern

LLM guardrails: screen every message in and out

Describe the hazards you care about, such as "is this a jailbreak attempt?", and score how harmful complying would be. Threshold the probabilities and your code decides whether to pass, review, block or reroute. TypeSafe's cookbook screens both directions of an LLM app with a single request.

LLM guardrails cookbook

RAG passage filtering

Score each retrieved passage with one request, then decide in code which ones reach the answering model. Keep and flag passages that contradict the question, and drop those carrying a hidden instruction or prompt injection.

RAG passages cookbook

Re-ranking and semantic search

In TypeSafe's re-ranking cookbook, one question per query and candidate pair lifted top-1 accuracy from 5% to 18% and top-10 accuracy from 38% to 62% on 40 legal queries, starting from BM25 shortlists of 30 passages. A separate cookbook builds line-by-line semantic search over a terms-of-service document.

Re-ranking cookbook

Citation checking

Catch wrong or hallucinated citations by checking a quote against its source. One Choice question decides whether the surrounding context supports the claim, and its confidence can flag the citation for human review.

Citation check cookbook

Structured extraction and function calling

Ask Jev for the parts named in a document, then resolve and validate them in code with confidence-based review. The same idea maps natural-language requests to ordinary typed functions: function names and closed-set arguments become confidence-aware questions.

Function calling cookbook

Real-time apps and large-scale data

At around 100 ms per call, AI can sit in code paths where UX is critical. TypeSafe's demos include a Doom bot making about 10 queries per second, which it puts at roughly $7 per hour, and a Wikipedia wikiracing agent. For big data, the pitch is turning huge corpora into features and insights with map-reduce style calls.

TypeSafe launch post