Skip to content

armgit / fuobe

Public Imported from GitHub
0

The open creative studio on ComfyUI: images, video, 3D and audio on open weights, on your own GPU or the cloud. Camera motions, effects, apps, Cinema, Layers, Image to 3D and an AI agent.

7.5 KB
Raw Blame
import "server-only";
import { cap, llmContext } from "./servers";
import type { ChatResult, LocalKind, LocalServer } from "./types";

/*
 * What went wrong with a local server, in words that say how to fix it: start the server, add a key,
 * load or download a model, or give it more context.
 */

/** How to get each server answering. */
export const START_HINT: Record<LocalKind, string> = {
  ollama: "Start it with `ollama serve` (or open the Ollama app).",
  lmstudio: "Start LM Studio's local server and load a model: in LM Studio, Developer → Start server (or run `lms server start`).",
  unsloth: "Start Unsloth Studio (the desktop app, or `unsloth studio -p 8888`) and load a model on its Chat page.",
  custom: "Start it, or check its URL in Settings → AI helper.",
};

/** Why a fetch failed: ECONNREFUSED, ENOTFOUND… */
function reason(e: unknown): string {
  const err = e as { message?: string; cause?: { code?: string; message?: string } } | undefined;
  return err?.cause?.code || err?.cause?.message || err?.message || "no answer";
}

export function unreachableError(s: LocalServer, e: unknown): Error {
  if ((e as Error | undefined)?.name === "TimeoutError") return new Error(`${cap(s.label)} took too long to answer. Try again, or pick a smaller model.`);
  return new Error(`Couldn't reach ${s.label} at ${s.url || "(no URL set)"} (${reason(e)}). ${START_HINT[s.kind]}`);
}

/**
 * The message inside an error body: OpenAI's {error: {message}}, {error: "…"}, FastAPI's {detail: "…"}
 * or {detail: {error: {message}}} (Unsloth Studio's context and pass-through errors), or the text.
 */
export function errorDetail(body: string): string {
  const text = (v: unknown): string | null => (typeof v === "string" && v ? v : null);
  type Err = { error?: string | { message?: unknown }; message?: unknown };
  const of = (j: Err | undefined): string | null => {
    const e = j?.error;
    return text(e) ?? (e && typeof e === "object" ? text(e.message) : null) ?? text(j?.message);
  };
  try {
    const j = JSON.parse(body) as Err & { detail?: unknown };
    const d = j?.detail;
    const msg = of(j) ?? text(d) ?? (d && typeof d === "object" ? of(d as Err) : null);
    if (msg) return msg.replace(/\s+/g, " ").trim().slice(0, 300);
  } catch {}
  return body.replace(/\s+/g, " ").trim().slice(0, 200);
}

/** A server that can't do structured output says so; the call is then made again without it. */
export const rejectsJson = (kind: LocalKind, status: number, body: string) =>
  status >= 400 && (/response_format|json_schema|json_object|grammar|structured output/i.test(body) || (kind === "ollama" && status === 400 && /\bformat\b/i.test(body)));

export function keyMessage(s: LocalServer): string {
  if (s.kind === "lmstudio") {
    return s.key
      ? "LM Studio rejected the API token. Check it in Settings → AI helper (LM Studio makes them under Developer → Server Settings → Manage Tokens)."
      : "LM Studio wants an API token (its Require Authentication setting is on): in LM Studio, Developer → Server Settings → Manage Tokens, then paste the token in Settings → AI helper.";
  }
  if (s.kind === "unsloth") {
    return s.key
      ? "Unsloth Studio rejected the API key. Create a new one in Unsloth Studio (avatar → Settings → API) and paste it in Settings → AI helper."
      : "Unsloth Studio needs an API key: in Unsloth Studio, avatar → Settings → API → Create, then paste the sk-unsloth-… key in Settings → AI helper.";
  }
  return s.key ? `${cap(s.label)} rejected the API key. Check it in Settings → AI helper.` : `${cap(s.label)} wants an API key. Add it in Settings → AI helper.`;
}

const CONTEXT = /context_length_exceeded|context (?:length|window|size)|n_ctx|tokens to keep|context the overflows|exceeds? (?:the )?(?:context|maximum)|maximum context|prompt is too long|too many tokens/i;
const NOT_LOADED = /no models? (?:is |are )?loaded|not loaded|no model is running/i;
const NO_MODEL = /model[^.]{0,80}(?:not found|does not exist|doesn't exist|unknown|no such|invalid)|unknown model|model_not_found|no such model/i;
const LOAD_FAILED = /failed to load|out of memory|insufficient (?:memory|vram)|not enough (?:memory|vram)|cuda error/i;

function contextMessage(s: LocalServer, model: string | null, detail: string): string {
  const n = /(?:context length of (?:only )?|n_ctx(?:_slot)?\s*[=:]?\s*)(\d[\d,]*)/i.exec(detail)?.[1];
  const who = model ? `${model}'s context` : "the model's context";
  const fix: Record<LocalKind, string> = {
    ollama: "Raise the context in Settings → AI helper (FUOBE_LLM_CONTEXT)",
    lmstudio: `Reload it with ${Math.round((llmContext() || 16384) / 1024)}k context (Settings → AI helper) or a larger one in LM Studio`,
    unsloth: "Load it in Unsloth Studio with a larger context",
    custom: "Start the server with a larger context",
  };
  return `This is longer than ${who}${n ? ` (${n} tokens)` : ""} on ${s.label}. ${fix[s.kind]}, or use shorter text.`;
}

function modelMessage(s: LocalServer, model: string | null, detail: string): string {
  const m = model ?? "that model";
  if (s.kind === "ollama") return `Ollama doesn't have ${m} yet — run \`ollama pull ${m}\``;
  if (s.kind === "unsloth") {
    return NOT_LOADED.test(detail) || !model
      ? "Unsloth Studio has no model loaded — load one on its Chat page (Select model), or turn on Model auto-switch (avatar → Settings → API) so the model picked in Fuobe loads by itself."
      : `Unsloth Studio has no model called ${m} — pick one it lists in Settings → AI helper, or leave the model empty to use the one loaded in Unsloth.`;
  }
  if (s.kind === "lmstudio") {
    return NOT_LOADED.test(detail) && !NO_MODEL.test(detail)
      ? `LM Studio has no model loaded — load one in LM Studio (or run \`lms load ${model ?? "<model>"}\`), or pick one in Settings → AI helper.`
      : `LM Studio has no model called ${m} — download it in LM Studio (or run \`lms get ${m}\`), or pick another in Settings → AI helper.`;
  }
  return `${cap(s.label)} has no model called ${m} — pick one it lists in Settings → AI helper.`;
}

/** A non-2xx answer from a chat call. */
export function httpError(s: LocalServer, status: number, body: string, model: string | null): Error {
  const detail = errorDetail(body);
  if (status === 401 || status === 403) return new Error(keyMessage(s));
  // The OpenAI code may be all that says so: {error: {message: "…", code: "context_length_exceeded"}}.
  if (CONTEXT.test(detail) || /"context_length_exceeded"/.test(body)) return new Error(contextMessage(s, model, detail));
  if (LOAD_FAILED.test(detail)) return new Error(`${cap(s.label)} couldn't load ${model ?? "the model"}: ${detail}. Free some GPU memory (ComfyUI shares it) or pick a smaller model.`);
  if (NOT_LOADED.test(detail) || NO_MODEL.test(detail) || (status === 404 && s.kind === "ollama")) return new Error(modelMessage(s, model, detail));
  if (status === 404) return new Error(`${cap(s.label)} has no chat endpoint at ${s.url} (404${detail ? `: ${detail}` : ""}). Check its URL in Settings → AI helper.`);
  return new Error(`${cap(s.label)} returned ${status}: ${detail}`);
}

/** A reply with no words in it. */
export function emptyReply(s: LocalServer, r: ChatResult): Error {
  if (r.stop === "length" && r.thought) return new Error(`${r.model} on ${s.label} spent its whole reply thinking. Turn thinking off for it, or pick a model that doesn't think first.`);
  if (r.stop === "length") return new Error(`${r.model} on ${s.label} ran out of room before it answered. Give it more context, or use shorter text.`);
  return new Error(`${cap(s.label)} returned no text`);
}