346 lines
11 KiB
TypeScript
346 lines
11 KiB
TypeScript
// extract_topology — Supabase Edge Function
|
|
// Reads a network-imports file from storage, sends it to Gemini with a strict
|
|
// JSON-schema prompt, and writes the parsed draft back to network_imports.
|
|
//
|
|
// Auth: caller must be admin or it_staff. Mirrors teams_crud auth pattern.
|
|
// Gemini model selection: 2.5 Flash by default; 2.5 Pro for PDFs >5 pages.
|
|
//
|
|
// Env vars required:
|
|
// SUPABASE_URL, SUPABASE_ANON_KEY, SUPABASE_SERVICE_ROLE_KEY, GEMINI_API_KEY
|
|
|
|
import { serve } from "https://deno.land/std@0.203.0/http/server.ts";
|
|
import { createClient } from "https://esm.sh/@supabase/supabase-js@2";
|
|
|
|
const CORS_HEADERS: Record<string, string> = {
|
|
"Access-Control-Allow-Origin": "*",
|
|
"Access-Control-Allow-Methods": "POST, OPTIONS",
|
|
"Access-Control-Allow-Headers":
|
|
"Content-Type, Authorization, apikey, x-client-info, x-requested-with, Origin, Accept",
|
|
"Access-Control-Max-Age": "3600",
|
|
};
|
|
|
|
const jsonResponse = (body: unknown, status = 200) =>
|
|
new Response(JSON.stringify(body), {
|
|
status,
|
|
headers: { "Content-Type": "application/json", ...CORS_HEADERS },
|
|
});
|
|
|
|
// Strict response schema we send to Gemini. Items get confidence scores so the
|
|
// review UI can rank "definitely a switch" above "might be a switch".
|
|
const RESPONSE_SCHEMA = {
|
|
type: "object",
|
|
properties: {
|
|
sites: {
|
|
type: "array",
|
|
items: {
|
|
type: "object",
|
|
properties: {
|
|
name: { type: "string" },
|
|
address: { type: "string" },
|
|
},
|
|
required: ["name"],
|
|
},
|
|
},
|
|
devices: {
|
|
type: "array",
|
|
items: {
|
|
type: "object",
|
|
properties: {
|
|
name: { type: "string" },
|
|
kind: {
|
|
type: "string",
|
|
enum: ["router", "switch", "ap", "firewall", "server", "endpoint", "patch_panel", "other"],
|
|
},
|
|
vendor: { type: "string" },
|
|
model: { type: "string" },
|
|
mgmt_ip: { type: "string" },
|
|
site_name: { type: "string" },
|
|
location_label: { type: "string" },
|
|
confidence: { type: "string", enum: ["high", "medium", "low"] },
|
|
},
|
|
required: ["name", "kind"],
|
|
},
|
|
},
|
|
ports: {
|
|
type: "array",
|
|
items: {
|
|
type: "object",
|
|
properties: {
|
|
device_name: { type: "string" },
|
|
port_number: { type: "string" },
|
|
port_kind: {
|
|
type: "string",
|
|
enum: ["rj45", "sfp", "sfp_plus", "console", "power", "wireless", "virtual", "other"],
|
|
},
|
|
access_vlan: { type: "integer" },
|
|
is_trunk: { type: "boolean" },
|
|
},
|
|
required: ["device_name", "port_number"],
|
|
},
|
|
},
|
|
links: {
|
|
type: "array",
|
|
items: {
|
|
type: "object",
|
|
properties: {
|
|
device_a: { type: "string" },
|
|
port_a: { type: "string" },
|
|
device_b: { type: "string" },
|
|
port_b: { type: "string" },
|
|
link_kind: {
|
|
type: "string",
|
|
enum: ["copper", "fiber", "wireless", "virtual", "unknown"],
|
|
},
|
|
confidence: { type: "string", enum: ["high", "medium", "low"] },
|
|
},
|
|
required: ["device_a", "device_b"],
|
|
},
|
|
},
|
|
vlans: {
|
|
type: "array",
|
|
items: {
|
|
type: "object",
|
|
properties: {
|
|
vlan_id: { type: "integer" },
|
|
name: { type: "string" },
|
|
description: { type: "string" },
|
|
},
|
|
required: ["vlan_id", "name"],
|
|
},
|
|
},
|
|
notes: { type: "string" },
|
|
},
|
|
};
|
|
|
|
const EXTRACTION_PROMPT = `You are a network-documentation reader. Look at the attached file (which may be a network diagram in PDF, an image of a diagram, a photo of a whiteboard, a Visio export, or a spreadsheet/notes) and extract every network device, port, link, VLAN, and site you can identify.
|
|
|
|
Rules:
|
|
- Output ONLY structured JSON conforming to the provided schema. No prose.
|
|
- For devices, classify kind precisely (switch / router / ap / etc.). If you genuinely cannot tell, use "other".
|
|
- For ports and links, give the exact labels as they appear ("Gi0/1", "Port 24", "uplink", etc.) — do not normalize.
|
|
- For each device and each link, attach a "confidence": "high" if the document clearly states it, "medium" if it's reasonably implied, "low" if it's a guess.
|
|
- If you cannot extract anything useful, return all arrays as empty and put a brief explanation in "notes".
|
|
- Do not invent items. Better to return less with high confidence than more with low.
|
|
- VLANs: only include if a VLAN ID number is visible.
|
|
- Sites: only include if a site/building name is clearly indicated.`;
|
|
|
|
interface ImportRow {
|
|
id: string;
|
|
storage_path: string;
|
|
file_kind: string;
|
|
status: string;
|
|
created_by: string;
|
|
}
|
|
|
|
function pickGeminiModel(fileKind: string, byteLength: number): string {
|
|
// Cheap default; escalate to Pro for big PDFs where layout reasoning matters.
|
|
if (fileKind === "pdf" && byteLength > 2 * 1024 * 1024) {
|
|
return "gemini-2.5-pro";
|
|
}
|
|
return "gemini-2.5-flash";
|
|
}
|
|
|
|
function mimeForFileKind(fileKind: string): string {
|
|
switch (fileKind) {
|
|
case "pdf":
|
|
return "application/pdf";
|
|
case "image":
|
|
// Default to PNG; Gemini accepts most image MIMEs and will sniff if needed.
|
|
return "image/png";
|
|
case "vsdx":
|
|
return "application/vnd.ms-visio.drawing";
|
|
case "csv":
|
|
return "text/csv";
|
|
case "xlsx":
|
|
return "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet";
|
|
case "docx":
|
|
return "application/vnd.openxmlformats-officedocument.wordprocessingml.document";
|
|
default:
|
|
return "application/octet-stream";
|
|
}
|
|
}
|
|
|
|
function bytesToBase64(bytes: Uint8Array): string {
|
|
// Deno has no btoa for binary; build base64 in chunks to avoid stack overflow.
|
|
let binary = "";
|
|
const chunkSize = 0x8000;
|
|
for (let i = 0; i < bytes.length; i += chunkSize) {
|
|
const slice = bytes.subarray(i, i + chunkSize);
|
|
binary += String.fromCharCode.apply(null, Array.from(slice));
|
|
}
|
|
return btoa(binary);
|
|
}
|
|
|
|
serve(async (req) => {
|
|
if (req.method === "OPTIONS") {
|
|
return new Response(null, { status: 204, headers: CORS_HEADERS });
|
|
}
|
|
if (req.method !== "POST") {
|
|
return jsonResponse({ error: "Method not allowed" }, 405);
|
|
}
|
|
|
|
const supabaseUrl = Deno.env.get("SUPABASE_URL") ?? "";
|
|
const anonKey = Deno.env.get("SUPABASE_ANON_KEY") ?? "";
|
|
const serviceKey = Deno.env.get("SUPABASE_SERVICE_ROLE_KEY") ?? "";
|
|
const geminiKey = Deno.env.get("GEMINI_API_KEY") ?? "";
|
|
if (!supabaseUrl || !anonKey || !serviceKey) {
|
|
return jsonResponse({ error: "Missing Supabase env config" }, 500);
|
|
}
|
|
if (!geminiKey) {
|
|
return jsonResponse({ error: "GEMINI_API_KEY not configured" }, 500);
|
|
}
|
|
|
|
// Auth caller
|
|
const authHeader = req.headers.get("Authorization") ?? "";
|
|
const token = authHeader.replace("Bearer ", "").trim();
|
|
if (!token) return jsonResponse({ error: "Missing access token" }, 401);
|
|
|
|
const authClient = createClient(supabaseUrl, anonKey, {
|
|
global: { headers: { Authorization: `Bearer ${token}` } },
|
|
});
|
|
const { data: authData, error: authError } = await authClient.auth.getUser();
|
|
if (authError || !authData?.user) return jsonResponse({ error: "Unauthorized" }, 401);
|
|
|
|
const adminClient = createClient(supabaseUrl, serviceKey);
|
|
|
|
// Caller must be admin or it_staff
|
|
const { data: profile, error: profileError } = await adminClient
|
|
.from("profiles")
|
|
.select("role")
|
|
.eq("id", authData.user.id)
|
|
.maybeSingle();
|
|
const role = (profile?.role ?? "").toString().toLowerCase();
|
|
if (profileError || (role !== "admin" && role !== "it_staff")) {
|
|
return jsonResponse({ error: "Forbidden" }, 403);
|
|
}
|
|
|
|
// Body: { import_id: string }
|
|
let body: { import_id?: string } = {};
|
|
try {
|
|
body = await req.json();
|
|
} catch (_e) {
|
|
return jsonResponse({ error: "Invalid JSON body" }, 400);
|
|
}
|
|
const importId = body.import_id;
|
|
if (!importId || typeof importId !== "string") {
|
|
return jsonResponse({ error: "Missing import_id" }, 400);
|
|
}
|
|
|
|
// Load the import row
|
|
const { data: imp, error: impErr } = await adminClient
|
|
.from("network_imports")
|
|
.select("id, storage_path, file_kind, status, created_by")
|
|
.eq("id", importId)
|
|
.maybeSingle<ImportRow>();
|
|
if (impErr || !imp) {
|
|
return jsonResponse({ error: "Import not found" }, 404);
|
|
}
|
|
if (imp.status === "applied" || imp.status === "discarded") {
|
|
return jsonResponse({ error: `Import already ${imp.status}` }, 409);
|
|
}
|
|
|
|
// Mark extracting
|
|
await adminClient
|
|
.from("network_imports")
|
|
.update({ status: "extracting", error_message: null })
|
|
.eq("id", importId);
|
|
|
|
try {
|
|
// Download from storage
|
|
const { data: fileData, error: dlErr } = await adminClient
|
|
.storage
|
|
.from("network-imports")
|
|
.download(imp.storage_path);
|
|
if (dlErr || !fileData) {
|
|
throw new Error(dlErr?.message ?? "Failed to download import file");
|
|
}
|
|
|
|
const arrayBuf = await fileData.arrayBuffer();
|
|
const bytes = new Uint8Array(arrayBuf);
|
|
const base64 = bytesToBase64(bytes);
|
|
const mime = mimeForFileKind(imp.file_kind);
|
|
const model = pickGeminiModel(imp.file_kind, bytes.length);
|
|
|
|
// Call Gemini
|
|
const geminiUrl =
|
|
`https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent?key=${geminiKey}`;
|
|
const geminiBody = {
|
|
contents: [
|
|
{
|
|
role: "user",
|
|
parts: [
|
|
{ text: EXTRACTION_PROMPT },
|
|
{ inlineData: { mimeType: mime, data: base64 } },
|
|
],
|
|
},
|
|
],
|
|
generationConfig: {
|
|
responseMimeType: "application/json",
|
|
responseSchema: RESPONSE_SCHEMA,
|
|
temperature: 0.1,
|
|
},
|
|
};
|
|
|
|
const geminiRes = await fetch(geminiUrl, {
|
|
method: "POST",
|
|
headers: { "Content-Type": "application/json" },
|
|
body: JSON.stringify(geminiBody),
|
|
});
|
|
|
|
if (!geminiRes.ok) {
|
|
const errBody = await geminiRes.text();
|
|
throw new Error(`Gemini ${geminiRes.status}: ${errBody.slice(0, 500)}`);
|
|
}
|
|
|
|
const geminiJson = await geminiRes.json();
|
|
const textOut: string =
|
|
geminiJson?.candidates?.[0]?.content?.parts?.[0]?.text ?? "";
|
|
|
|
let parsed: unknown;
|
|
try {
|
|
parsed = JSON.parse(textOut);
|
|
} catch (_e) {
|
|
throw new Error("Gemini returned non-JSON output");
|
|
}
|
|
|
|
// Minimum-quality gate: must have at least one device, or status=failed.
|
|
const deviceCount = Array.isArray((parsed as { devices?: unknown[] })?.devices)
|
|
? (parsed as { devices: unknown[] }).devices.length
|
|
: 0;
|
|
|
|
if (deviceCount === 0) {
|
|
await adminClient
|
|
.from("network_imports")
|
|
.update({
|
|
status: "failed",
|
|
ai_result: parsed as Record<string, unknown>,
|
|
error_message: "No devices extracted from document. Start from blank canvas.",
|
|
})
|
|
.eq("id", importId);
|
|
return jsonResponse(
|
|
{ status: "failed", reason: "no_devices", ai_result: parsed },
|
|
200,
|
|
);
|
|
}
|
|
|
|
// Store + mark for review
|
|
await adminClient
|
|
.from("network_imports")
|
|
.update({
|
|
status: "review",
|
|
ai_result: parsed as Record<string, unknown>,
|
|
error_message: null,
|
|
})
|
|
.eq("id", importId);
|
|
|
|
return jsonResponse({ status: "review", ai_result: parsed });
|
|
} catch (err) {
|
|
const msg = (err as Error).message ?? String(err);
|
|
await adminClient
|
|
.from("network_imports")
|
|
.update({ status: "failed", error_message: msg })
|
|
.eq("id", importId);
|
|
return jsonResponse({ error: msg }, 500);
|
|
}
|
|
});
|