import { LlamaModel, LlamaContext, LlamaChatSession } from "node-llama-cpp";
import FHIR from "fhir-kit-client";
import sqlite3 from "sqlite3";
import path from "path";
import fs from "fs";
import { fileURLToPath } from "url";
const __filename = fileURLToPath(import.meta.url);
const __dirname = path.dirname(__filename);
// ---- Resolve model path (env > explicit file > first .gguf in models/) ----
function resolveModelPath(): string {
if (process.env.MODEL_PATH && fs.existsSync(process.env.MODEL_PATH)) {
return process.env.MODEL_PATH;
}
const modelsDir = path.join(__dirname, "..", "models");
if (!fs.existsSync(modelsDir)) {
throw new Error(`Models directory not found at: ${modelsDir}`);
}
const ggufs = fs.readdirSync(modelsDir).filter(f => f.endsWith(".gguf"));
if (ggufs.length === 0) {
throw new Error(`No .gguf model found in ${modelsDir}. Place your model there or set MODEL_PATH.`);
}
// Prefer TinyLlama if present, else first gguf
const tiny = ggufs.find(f => f.toLowerCase().includes("tinyllama")) || ggufs[0];
return path.join(modelsDir, tiny);
}
const MODEL_PATH = resolveModelPath();
// ---- SQLite (demo DB name: mimiciii_demo.db) ----
const dbPath = path.join(__dirname, "..", "mimiciii_demo.db");
const db = new sqlite3.Database(dbPath, (err) => {
if (err) console.warn("⚠️ SQLite open error:", err.message);
else console.log(`✅ SQLite DB opened: ${dbPath}`);
});
// ---- FHIR client to HAPI public server ----
const fhirClient = new FHIR({ baseUrl: "https://hapi.fhir.org/baseR4" });
// ---- Initialize local LLM session ----
let session: LlamaChatSession;
try {
const model = new LlamaModel({ modelPath: MODEL_PATH });
const context = new LlamaContext({ model, contextSize: 2048 }); // << nCtx -> contextSize
session = new LlamaChatSession({ context });
console.log(`✅ Local LLM initialized: ${MODEL_PATH}`);
} catch (e) {
console.error("❌ Failed to initialize local GGUF model:", e);
throw e;
}
export async function getSummaryFromDB({ patientId }: { patientId: string }): Promise<any> {
return new Promise((resolve) => {
const sql = "SELECT TEXT FROM NOTEEVENTS WHERE SUBJECT_ID = ? AND CATEGORY = 'Discharge summary' LIMIT 1";
db.get(sql, [patientId], async (err, row: { TEXT: string } | undefined) => {
if (err || !row) {
return resolve({ summary: `No discharge summary found for patient ID ${patientId} in the local DB.` });
}
try {
console.log(`[Local LLM] Summarizing note from DB for patient ${patientId}...`);
const prompt =
"You are a clinical AI assistant. Summarize the following discharge note into a concise, factual paragraph:\n\n" +
row.TEXT;
const result = await session.prompt(prompt, { maxTokens: 160, temperature: 0.3 });
resolve({ source: "SQLite (MIMIC-III)", patientId, ai_summary: result });
} catch (e) {
resolve({ summary: "Failed to generate AI summary from DB note." });
}
});
});
}
export async function getSummaryFromFHIR({ patientId }: { patientId: string }): Promise<any> {
try {
console.log(`[FHIR] Reading patient: ${patientId}`);
const patient: any = await fhirClient.read({ resourceType: "Patient", id: patientId }).catch(() => null);
if (!patient) return { summary: `Patient "${patientId}" not found.` };
const conditions: any = await fhirClient.search({
resourceType: "Condition",
searchParams: { patient: patientId, clinical_status: "active" }
}).catch(() => null);
const observations: any = await fhirClient.search({
resourceType: "Observation",
searchParams: { patient: patientId, _sort: "-date", _count: "5" }
}).catch(() => null);
const given = patient?.name?.[0]?.given?.join(" ") || "Unknown";
const family = patient?.name?.[0]?.family || "";
const birth = patient?.birthDate || "Unknown";
let clinicalText = `Patient: ${given} ${family}, born ${birth}.\n\nActive Conditions:\n`;
const condEntries = conditions?.entry ?? [];
if (condEntries.length) {
for (const entry of condEntries) {
const text = entry?.resource?.code?.text || entry?.resource?.code?.coding?.[0]?.display || "Condition";
clinicalText += `- ${text}\n`;
}
} else {
clinicalText += "- No active conditions found.\n";
}
clinicalText += "\nRecent Observations:\n";
const obsEntries = observations?.entry ?? [];
if (obsEntries.length) {
for (const entry of obsEntries) {
const obs = entry?.resource;
const label = obs?.code?.text || obs?.code?.coding?.[0]?.display || "Observation";
if (obs?.valueQuantity?.value != null) {
const val = parseFloat(String(obs.valueQuantity.value));
const unit = obs?.valueQuantity?.unit || "";
const dt = obs?.effectiveDateTime || "";
clinicalText += `- ${label}: ${isNaN(val) ? obs.valueQuantity.value : val.toFixed(2)} ${unit} ${dt ? `on ${dt}` : ""}\n`;
} else if (obs?.valueCodeableConcept?.text) {
clinicalText += `- ${label}: ${obs.valueCodeableConcept.text}\n`;
} else {
clinicalText += `- ${label}\n`;
}
}
} else {
clinicalText += "- No recent observations found.\n";
}
console.log(`[Local LLM] Summarizing FHIR data for patient ${patientId}...`);
const prompt =
"You are a clinical AI assistant. Summarize the following patient record into a concise, single paragraph:\n\n" +
clinicalText;
const result = await session.prompt(prompt, { maxTokens: 160, temperature: 0.3 });
return {
source: "FHIR",
patientId: patient.id,
patientName: `${given} ${family}`.trim(),
ai_summary: result
};
} catch (error) {
return { summary: `An error occurred while summarizing FHIR data for patient "${patientId}".` };
}
}
export async function searchGuidelines({ topic }: { topic: string }): Promise<any> {
// Placeholder. Replace with PubMed/NICE/etc. later.
const mockGuidelines = [
{
guidelineId: "GUID-HTN-01",
topic: "hypertension",
title: "2024 ACC/AHA Guideline for the Management of Hypertension"
}
];
const searchTopic = topic.toLowerCase();
const results = mockGuidelines.filter((g) => g.topic.includes(searchTopic));
return results.length > 0 ? results : { error: `No guidelines found for topic "${topic}".` };
}
