ollama-plus/rag-server/index.ts

import { serve } from "bun";
import fs from "node:fs";
import path from "node:path";

// types
interface Chunk {
	id: string;
	text: string;
	metadata?: Record<string, unknown>;
	vector: number[];
}

interface Collection {
	name: string;
	chunks: Chunk[];
}

interface OllamaChatMessage {
	role: "system" | "user" | "assistant";
	content: string;
}

interface OllamaChatRequest {
	model?: string;
	messages: OllamaChatMessage[];
	stream?: boolean;
}

interface OllamaChatResponse {
	message?: OllamaChatMessage;
	[k: string]: unknown;
}

interface UpsertInputItem {
	text: string;
	metadata?: Record<string, unknown>;
}

interface OpenAPIObject {
	openapi: string;
	info: { title: string; version: string };
	paths: Record<string, unknown>;
}

// env
const PORT: number = Number(process.env.PORT || 8788),
	HOST: string = process.env.HOST || "0.0.0.0",
	OLLAMA_BASE: string = process.env.OLLAMA_BASE || "http://localhost:11434",
	OLLAMA_CHAT_MODEL: string = process.env.OLLAMA_CHAT_MODEL || "llama3.1",
	OLLAMA_EMBED_MODEL: string = process.env.OLLAMA_EMBED_MODEL || "nomic-embed-text",
	DATA_DIR: string = process.env.DATA_DIR || path.resolve("./data"),
	SNAPSHOT: string = path.join(DATA_DIR, "rag.json");

// in-memory db
const db: Map<string, Collection> = new Map();

// util: smol json persistence
function ensureDirs(): void {
	if (!fs.existsSync(DATA_DIR)) fs.mkdirSync(DATA_DIR, { recursive: true });
}

// you can probably guess
function loadSnapshot(): void {
	try {
		ensureDirs();
		if (fs.existsSync(SNAPSHOT)) {
			const raw = fs.readFileSync(SNAPSHOT, "utf8");
			const obj = JSON.parse(raw || "{}") as Record<string, Collection>;
			for (const [name, value] of Object.entries(obj)) db.set(name, value);
		}
	} catch (e) {
		console.warn("failed to load snapshot:", e);
	}
}

// you can probably guess 2
function saveSnapshot(): void {
	try {
		ensureDirs();
		const obj = Object.fromEntries(db.entries());
		fs.writeFileSync(SNAPSHOT, JSON.stringify(obj, null, 2));
	} catch (e) {
		console.warn("failed to save snapshot:", e);
	}
}

loadSnapshot();

// basic text splitter (recursive by punctuation, then by length)
function chunkText(text: string, maxLen = 800): string[] {
	const parts = text
		.split(/\n{2,}/g)
		.flatMap(p => p.split(/(?<=[.!?])\s+/g))
		.flatMap(s => s.length > maxLen ? s.match(new RegExp(`.{1,${maxLen}}`, "g")) || [] : [s])
		.map(s => s.trim())
		.filter(Boolean);
	return parts;
}

// cosine similarity
function dot(a: number[], b: number[]): number { let s = 0; for (let i = 0; i < a.length; i++) s += (a[i] || 0) * (b[i] || 0); return s; }
function norm(a: number[]): number { return Math.sqrt(dot(a, a)); }
function cosineSim(a: number[], b: number[]): number { const d = dot(a, b), n = norm(a) * norm(b) || 1; return d / n; }

// call ollama embeddings
async function embedAll(texts: string[]): Promise<number[][]> {
	const primary = await fetch(`${OLLAMA_BASE}/api/embed`, {
		method: "POST",
		headers: { "content-type": "application/json" },
		body: JSON.stringify({ model: OLLAMA_EMBED_MODEL, input: texts })
	});

	if (primary.ok) {
		const j: { embeddings: number[][] } = await primary.json();
		return j.embeddings;
	}

	const results: number[][] = [];
	for (const t of texts) {
		const r = await fetch(`${OLLAMA_BASE}/api/embeddings`, {
			method: "POST",
			headers: { "content-type": "application/json" },
			body: JSON.stringify({ model: OLLAMA_EMBED_MODEL, prompt: t })
		});

		if (!r.ok) throw new Error(`embed failed: ${r.status}`);

		const j: { embedding: number[] } = await r.json();
		results.push(j.embedding);
	}
	return results;
}

// call ollama chat/generate with retrieved context
async function ollamaChat(req: OllamaChatRequest): Promise<OllamaChatResponse> {
	const res = await fetch(`${OLLAMA_BASE}/api/chat`, {
		method: "POST",
		headers: { "content-type": "application/json" },
		body: JSON.stringify({ model: req.model || OLLAMA_CHAT_MODEL, messages: req.messages, stream: req.stream })
	});

	if (!res.ok) throw new Error(`ollama chat failed: ${res.status}`);
	const j: OllamaChatResponse = await res.json();

	return j;
}

// openapi for open webui tool integration
const OPENAPI: OpenAPIObject = {
	openapi: "3.1.0",
	info: { title: "RAG Server (Ollama)", version: "1.0.0" },
	paths: {
		"/collections": {
			get: { operationId: "listCollections" },
			post: { operationId: "createCollection" }
		},
		"/upsert": { post: { operationId: "upsert" } },
		"/query": { post: { operationId: "query" } },
		"/chat": { post: { operationId: "chat" } }
	}
};

// tiny router
async function json<T = any>(req: Request): Promise<T> { try { return await req.json() as T; } catch { return {} as T; } }
function sendJson(_res: unknown, status: number, obj: unknown): Response {
	return new Response(JSON.stringify(obj), { status, headers: { "content-type": "application/json; charset=utf-8" } });
}

async function handleCollections(req: Request): Promise<Response> {
	if (req.method === "GET") {
		return sendJson(null, 200, { collections: Array.from(db.keys()) });
	}

	if (req.method === "POST") {
		const body = await json<{ name?: string }>(req),
			name = String(body?.name || "").trim();

		if (!name) return sendJson(null, 400, { error: "name required" });
		if (!db.has(name)) db.set(name, { name, chunks: [] });

		saveSnapshot();
		return sendJson(null, 200, { ok: true });
	}

	return new Response("not found", { status: 404 });
}

async function handleUpsert(req: Request): Promise<Response> {
	const body = await json<{ collection?: string; items?: UpsertInputItem[] }>(req),
		collection = String(body?.collection || "").trim(),
		items: UpsertInputItem[] = Array.isArray(body?.items) ? body.items : [];

	if (!collection) return sendJson(null, 400, { error: "collection required" });
	if (!db.has(collection)) db.set(collection, { name: collection, chunks: [] });

	const col = db.get(collection)!,
		chunksToIndex: { text: string; metadata?: Record<string, unknown>; _id: string }[] = [];

	for (const it of items) {
		const parts = chunkText(String(it.text || ""));
		for (const p of parts) chunksToIndex.push({ text: p, metadata: it.metadata || {}, _id: crypto.randomUUID() });
	}

	const vecs = await embedAll(chunksToIndex.map(x => x.text));
	for (let i = 0; i < chunksToIndex.length; i++) {
		const item = chunksToIndex[i],
			doc: Chunk = { id: item._id, text: item.text, metadata: item.metadata, vector: vecs[i] };

		col.chunks.push(doc);
	}

	saveSnapshot();
	return sendJson(null, 200, { ok: true, indexed: chunksToIndex.length });
}

async function handleQuery(req: Request): Promise<Response> {
	const body = await json<{ collection?: string; query?: string; topK?: number }>(req),
		collection = String(body?.collection || "").trim(),
		query = String(body?.query || "").trim(),
		topK = Number(body?.topK || 5);

	if (!collection || !query) return sendJson(null, 400, { error: "collection and query required" });

	const col = db.get(collection);
	if (!col) return sendJson(null, 404, { error: "collection not found" });

	const [qvec] = await embedAll([query]),
		scored = col.chunks.map((c) => ({ c, score: cosineSim(qvec, c.vector) }))
			.sort((a, b) => b.score - a.score)
			.slice(0, topK)
			.map(x => ({ id: x.c.id, text: x.c.text, metadata: x.c.metadata, score: x.score }));
	return sendJson(null, 200, { matches: scored });
}

async function handleChat(req: Request): Promise<Response> {
	const body = await json<{ collection?: string; query?: string; topK?: number; model?: string }>(req),
		collection = String(body?.collection || "").trim(),
		query = String(body?.query || "").trim(),
		topK = Number(body?.topK || 5),
		model = body?.model || OLLAMA_CHAT_MODEL;

	if (!collection || !query) return sendJson(null, 400, { error: "collection and query required" });

	const col = db.get(collection);
	if (!col) return sendJson(null, 404, { error: "collection not found" });

	const [qvec] = await embedAll([query]),
		matches = col.chunks.map((c) => ({ c, score: cosineSim(qvec, c.vector) }))
			.sort((a, b) => b.score - a.score)
			.slice(0, topK);

	const context = matches.map((m, i) => `[[doc ${i + 1} score=${m.score.toFixed(3)}]]\n${m.c.text}`).join("\n\n"),
		system: string = `you are a helpful assistant. use ONLY the provided context to answer. if the answer isn't in the context, say you don't know. cite as [doc N].`,
		user: string = `question: ${query}\n\ncontext:\n${context}`;

	const out = await ollamaChat({ model, messages: [{ role: "system", content: system }, { role: "user", content: user }], stream: false });
	return sendJson(null, 200, {
		answer: out?.message?.content || "",
		citations: matches.map((m, i) => ({ id: m.c.id, score: m.score, text: m.c.text }))
	});
}

const pickFunc = (pathname: string) => {
	switch (pathname) {
		case "/collections":
			return handleCollections;
		case "/upsert":
			return handleUpsert;
		case "/query":
			return handleQuery;
		case "/chat":
			return handleChat;
		default:
			return undefined;
	}
}

const server = serve({
	port: PORT,
	hostname: HOST,
	fetch: async (req: Request): Promise<Response> => {
		const u = new URL(req.url);
		if (req.method === "GET" && u.pathname === "/") return new Response("ok");
		if (req.method === "GET" && u.pathname === "/openapi.json") return sendJson(null, 200, OPENAPI);
		return pickFunc(u.pathname)?.call(req) || new Response("not found", { status: 404 });
	}
});

console.log(`[rag] listening on http://${HOST}:${PORT}`);