@lenml/jevseek

DeepSeek FIM and llama.cpp completion adapter for Jev-style SystemOne decisions.

pnpm add @lenml/jevseek

import { createJevSeek } from "@lenml/jevseek";

const client = createJevSeek({
  apiKey: process.env.DEEPSEEK_API_KEY,
  model: "deepseek-flash",
});

const result = await client.systemOne({
  state: "I was charged twice and need help today.",
  questions: {
    department: {
      type: "choice",
      instructions: "Which team should handle this?",
      criteria: {
        billing: "Payment or subscription",
        technical: "Bug or integration",
        sales: "Pricing or account",
      },
    },
    urgent: {
      type: "noul",
      instructions: "The request is time-sensitive.",
    },
  },
});

Question types:

• choice
• score
• noul

Features:

• Custom or default prompt templates, with a strict fallback for missing candidate logprobs.
• DeepSeek Beta FIM transport.
• llama.cpp native /completion transport and multimodal_data.
• Candidate logprob normalization.
• Jev-shaped choice, score, and noul answers.
• Per-question concurrency limits.
• Timeout, AbortSignal, exponential retry, and Retry-After.
• Injectable fetch and transport.
• Debug diagnostics for usage, prompts, probabilities, and request IDs.

The default DeepSeek base URL is https://api.deepseek.com/beta. Runs on Node.js 18+, Bun, workers, and browsers.

Provider-specific defaults are used when promptTemplate is not set. DeepSeek uses a function-completion template that lands the first token on a candidate code. llama.cpp keeps the readable Answer: \boxed{ template.

Cost estimates

For a 200-token state and three 100-token questions, JevSeek sends three DeepSeek FIM requests and repeats the state: about 900 input tokens and 3 output tokens. Jev processes the same state and questions in one input-only request: about 500 tokens. Prices below use the 2026-09-22 published rates and no cache, retry, or concurrency discount.

| Backend | 1,000 SystemOne requests | Relative to Jev | | ------------------- | ------------------------ | --------------- | | Jev | ¥0.15 | 1.0× | | DeepSeek Flash idle | ¥0.91 | 6.0× | | DeepSeek Flash peak | ¥1.82 | 12.1× | | DeepSeek Pro idle | ¥4.09 | 27.1× | | DeepSeek Pro peak | ¥8.18 | 54.1× |

The comparison uses $1 = ¥7.2 for Jev's USD price. DeepSeek prompt-cache hits, additional questions, fallback retries, and longer states change the result. See the full cost model for formulas and llama.cpp estimates for 2B, 4B, 9B, 28B, and 30B-A3B models.

llama.cpp

const local = createJevSeek({
  provider: "llamacpp",
  baseUrl: "http://127.0.0.1:8080/v1",
  model: "local-model",
});

const result = await local.systemOne({
  state: "The request is urgent.",
  questions: {
    urgent: {
      type: "noul",
      instructions: "Is it urgent?",
    },
  },
  multimodal_data: ["<base64-data>"],
});

llama.cpp mode sends POST /completion with n_probs. A /v1 base URL suffix is removed because /completion is a native server route. multimodal_data is accepted only in this mode; DeepSeek mode rejects it before sending a request. The model needs a matching multimodal projector, and the prompt needs one server media marker per data entry.

The default llama.cpp base URL is http://127.0.0.1:8080/v1.

Override promptTemplate in createJevSeek() for the client, or in systemOne() for one request. String templates support {{state}}, {{question}}, {{instructions}}, {{options}}, {{questionType}}, and {{codes}}; function templates receive the structured rendering context.

If the response contains no candidate logprob, the client retries with the provider-specific fallback template. Use fallbackPromptTemplate to override it at client or request level. If that still fails, the default missingLogprobPolicy: "zero" returns zero probabilities with confidence 0; set it to "error" to retain the parse error.

The DeepSeek defaults were measured on all 70 MMLU-Pro validation rows on 2026-09-23. deepseek-flash scored 74-76% across repeat runs and deepseek-v4-pro scored 78.57%, with no zero-probability fallback. DeepSeek FIM currently returns a usable logprob only for the sampled token; other top_logprobs are commonly -9999, so results are one-hot selections rather than calibrated distributions.

Run pnpm prompt:benchmark to reproduce the llama.cpp JevBench Easy prompt score.

See the repository documentation for the full HTTP and browser API.