297 lines
10 KiB
JavaScript
297 lines
10 KiB
JavaScript
import { z } from "zod";
|
|
import { reconstructionV2Schema } from "../reconstruction/schema.js";
|
|
|
|
/**
|
|
* Provider abstraction — the app calls getProvider() which returns an object
|
|
* with a generateReconstruction(scenario, modelName) method.
|
|
* Only Ollama is implemented right now; swapping providers requires only
|
|
* changing getProvider().
|
|
*/
|
|
|
|
export function getProvider() {
|
|
return new OllamaLlmProvider();
|
|
}
|
|
|
|
function recoverJson(raw) {
|
|
if (typeof raw !== "string") return raw;
|
|
const trimmed = raw.trim();
|
|
|
|
if (trimmed.length === 0) throw new SyntaxError("Model produced empty output");
|
|
|
|
try {
|
|
return JSON.parse(trimmed);
|
|
} catch {
|
|
// Not directly parseable — try closing braces/brackets from the right side
|
|
}
|
|
|
|
let result = trimmed;
|
|
|
|
let braceDepth = 0;
|
|
let bracketDepth = 0;
|
|
let inString = false;
|
|
let escaped = false;
|
|
|
|
for (let i = 0; i < result.length; i++) {
|
|
const ch = result[i];
|
|
|
|
if (escaped) { escaped = false; continue; }
|
|
if (ch === '\\') { escaped = true; continue; }
|
|
if (ch === '"') { inString = !inString; continue; }
|
|
if (inString) continue;
|
|
if (ch === '{') braceDepth++;
|
|
else if (ch === '}') braceDepth--;
|
|
else if (ch === '[') bracketDepth++;
|
|
else if (ch === ']') bracketDepth--;
|
|
}
|
|
|
|
const closingBrackets = [];
|
|
for (let i = 0; i < bracketDepth; i++) closingBrackets.push(']');
|
|
for (let i = 0; i < braceDepth; i++) closingBrackets.push('}');
|
|
|
|
if (closingBrackets.length > 0) {
|
|
const closed = result + closingBrackets.reverse().join('');
|
|
try { return JSON.parse(closed); } catch { /* still broken */ }
|
|
}
|
|
|
|
const lastOpen = Math.max(result.lastIndexOf('{'), result.lastIndexOf('['));
|
|
if (lastOpen >= 0) {
|
|
try { return JSON.parse(result.slice(lastOpen)); } catch { /* nothing works */ }
|
|
}
|
|
|
|
throw new SyntaxError("Model output could not be parsed as JSON: " + result.slice(0, 300) + "...");
|
|
}
|
|
|
|
let _chatSupported = null;
|
|
const reconstructionJsonSchema = z.toJSONSchema(reconstructionV2Schema);
|
|
|
|
/** @internal Test-only seam for isolated provider capability scenarios. */
|
|
export function __resetChatSupportForTests() {
|
|
_chatSupported = null;
|
|
}
|
|
|
|
async function detectChatSupport(baseUrl, modelName) {
|
|
if (_chatSupported !== null) return _chatSupported;
|
|
|
|
try {
|
|
const res = await fetch(`${baseUrl}/api/chat`, {
|
|
method: "POST",
|
|
headers: { "Content-Type": "application/json" },
|
|
body: JSON.stringify({
|
|
model: modelName,
|
|
messages: [{ role: "user", content: "test" }],
|
|
stream: false,
|
|
}),
|
|
});
|
|
|
|
if (res.ok) {
|
|
await res.text();
|
|
_chatSupported = true;
|
|
} else if (res.status === 405 || res.status === 501) {
|
|
await res.text();
|
|
_chatSupported = false;
|
|
} else {
|
|
await res.text();
|
|
_chatSupported = false;
|
|
}
|
|
} catch {
|
|
_chatSupported = false;
|
|
}
|
|
|
|
return _chatSupported;
|
|
}
|
|
|
|
class OllamaLlmProvider {
|
|
async generateReconstruction(scenario, modelName) {
|
|
// scenario is ALREADY a fully-built prompt text (built by analyseScenario).
|
|
// Do NOT call buildPrompt() again — that would double-wrap the prompt.
|
|
const prompt = scenario;
|
|
|
|
const baseUrl = process.env.OLLAMA_BASE_URL;
|
|
if (!baseUrl) throw new Error("OLLAMA_BASE_URL is not set");
|
|
|
|
let apiUsed = null;
|
|
let chatSupported = false;
|
|
let rawResponse = null;
|
|
let fullResponseData = null;
|
|
const providerExecution = {
|
|
chatCapabilityDetected: false,
|
|
chatRequestAttempted: false,
|
|
chatRequestSucceeded: false,
|
|
generateRequestAttempted: false,
|
|
};
|
|
|
|
// ================================================================
|
|
// Step 1: Detect whether /api/chat exists (cache result)
|
|
// ================================================================
|
|
try {
|
|
chatSupported = await detectChatSupport(baseUrl, modelName);
|
|
} catch { /* failed silently — defaults to false */ }
|
|
providerExecution.chatCapabilityDetected = chatSupported;
|
|
|
|
// ================================================================
|
|
// Step 2: Try /api/chat if supported with the reconstruction schema
|
|
// ================================================================
|
|
if (chatSupported) {
|
|
try {
|
|
const controller = new AbortController();
|
|
const timeout = setTimeout(() => controller.abort(), 60000);
|
|
|
|
apiUsed = "/api/chat";
|
|
providerExecution.chatRequestAttempted = true;
|
|
const res = await fetch(`${baseUrl}/api/chat`, {
|
|
method: "POST",
|
|
headers: { "Content-Type": "application/json" },
|
|
body: JSON.stringify({
|
|
model: modelName,
|
|
messages: [{ role: "user", content: prompt }],
|
|
stream: false,
|
|
format: reconstructionJsonSchema,
|
|
}),
|
|
signal: controller.signal,
|
|
});
|
|
|
|
clearTimeout(timeout);
|
|
|
|
if (res.ok) {
|
|
providerExecution.chatRequestSucceeded = true;
|
|
fullResponseData = await res.json();
|
|
rawResponse = typeof fullResponseData.message?.content === "string"
|
|
? fullResponseData.message.content
|
|
: JSON.stringify(fullResponseData.message?.content ?? null);
|
|
apiUsed = "/api/chat";
|
|
} else {
|
|
await res.text();
|
|
}
|
|
} catch (e) {
|
|
if (!e.message.includes("abort")) { /* non-fatal */ }
|
|
}
|
|
}
|
|
|
|
// ================================================================
|
|
// Step 3: /api/generate (works on all Ollama versions)
|
|
// Use a long timeout — cold starts can take 2-4 minutes for large models.
|
|
// ================================================================
|
|
if (rawResponse == null || rawResponse === "") {
|
|
try {
|
|
const controller = new AbortController();
|
|
const timeout = setTimeout(() => controller.abort(), 300000); // 5 min for cold start
|
|
|
|
apiUsed = "/api/generate"; // set BEFORE the request so we know which API failed
|
|
providerExecution.generateRequestAttempted = true;
|
|
|
|
const res = await fetch(`${baseUrl}/api/generate`, {
|
|
method: "POST",
|
|
headers: { "Content-Type": "application/json" },
|
|
body: JSON.stringify({
|
|
model: modelName,
|
|
prompt,
|
|
stream: false,
|
|
// No format:json — older Ollama doesn't support it on /api/generate either.
|
|
// We rely on the strong prompt instruction above instead.
|
|
}),
|
|
signal: controller.signal,
|
|
});
|
|
|
|
clearTimeout(timeout);
|
|
|
|
if (!res.ok) {
|
|
const text = await res.text().catch(() => "");
|
|
throw new Error(`/api/generate returned ${res.status}: ${text.slice(0, 500)}`);
|
|
}
|
|
|
|
fullResponseData = await res.json();
|
|
|
|
rawResponse = typeof fullResponseData.response === "string"
|
|
? fullResponseData.response
|
|
: JSON.stringify(fullResponseData);
|
|
|
|
} catch (e) {
|
|
if (apiUsed === "/api/generate") {
|
|
const error = new Error(
|
|
`Ollama /api/generate request timed out after 5 minutes.\n\n` +
|
|
`This usually means:\n` +
|
|
`1. The model is loading into memory for the first time (cold start) — this can take several minutes\n` +
|
|
`2. Your hardware is slow for this model size\n` +
|
|
`3. Ollama server is overloaded\n\n` +
|
|
`Try:\n` +
|
|
`- Run the request again after ~1 minute (model may be cached now)\n` +
|
|
`- Use a smaller model (e.g., llama3.1 instead of llama3.1:70b)\n` +
|
|
`- Check Ollama logs: \`ollama serve\` or look at your system logs`
|
|
);
|
|
error.providerApiPath = apiUsed;
|
|
error.providerExecution = providerExecution;
|
|
throw error;
|
|
}
|
|
throw e;
|
|
}
|
|
}
|
|
|
|
// ================================================================
|
|
// Step 4: Diagnose empty output
|
|
// ================================================================
|
|
if (rawResponse === "") {
|
|
let diagInfo = "";
|
|
|
|
if (fullResponseData) {
|
|
const keys = Object.keys(fullResponseData);
|
|
diagInfo = "Keys in response: " + keys.join(", ") + "\n";
|
|
|
|
for (const key of keys) {
|
|
const val = fullResponseData[key];
|
|
if (typeof val === "string") {
|
|
diagInfo += ` ${key}: "${val.slice(0, 200)}"\n`;
|
|
} else if (typeof val === "object" && val != null) {
|
|
try { diagInfo += ` ${key}: ${JSON.stringify(val).slice(0, 300)}\n`; } catch { diagInfo += ` ${key}: [object]\n`; }
|
|
} else {
|
|
diagInfo += ` ${key}: ${String(val)}\n`;
|
|
}
|
|
}
|
|
}
|
|
|
|
const error = new Error(
|
|
"Model produced empty output.\n\n" +
|
|
"API used: " + (apiUsed || "none") + "\n" +
|
|
"/api/chat supported: " + chatSupported + "\n" +
|
|
"Full server response:\n" + (diagInfo || "(none)\n") +
|
|
"\nPossible causes:\n" +
|
|
"- Check 'ollama list' — make sure the model name matches exactly what's installed\n" +
|
|
"- The model may be corrupted. Try: ollama pull " + modelName + "\n" +
|
|
"- This Ollama version does not support format:json — using prompt instructions only (reliability varies)\n" +
|
|
"- If your model is very small (e.g., tinyllama, phi), try a larger one like llama3.1 or mistral"
|
|
);
|
|
error.providerApiPath = apiUsed;
|
|
error.providerExecution = providerExecution;
|
|
throw error;
|
|
}
|
|
|
|
// ================================================================
|
|
// Step 5: Parse and return
|
|
// ================================================================
|
|
try {
|
|
return {
|
|
response: recoverJson(rawResponse),
|
|
providerApiPath: apiUsed,
|
|
providerExecution,
|
|
};
|
|
} catch (e) {
|
|
if (e instanceof SyntaxError) {
|
|
const error = new Error(
|
|
"Model returned output that could not be parsed as valid JSON.\n\n" +
|
|
"API used: " + (apiUsed || "none") + "\n" +
|
|
"/api/chat supported: " + chatSupported + "\n" +
|
|
"Raw model output:\n" + rawResponse.slice(0, 1000) + (rawResponse.length > 1000 ? "\n...(truncated)" : "") +
|
|
"\n\nPossible causes:\n" +
|
|
"- This Ollama version does not support format:json. The model is producing free-form text.\n" +
|
|
"- Try a larger model (llama3.1, mistral-large) which follows JSON instructions better\n" +
|
|
"- Shorten your scenario to under 500 words\n" +
|
|
"- Consider upgrading Ollama: https://ollama.com/download"
|
|
);
|
|
error.providerApiPath = apiUsed;
|
|
error.providerExecution = providerExecution;
|
|
throw error;
|
|
}
|
|
throw e;
|
|
}
|
|
}
|
|
}
|