Roadbound-BRR/supabase/functions/operations-stop-alias-enhance/index.ts

import { fail, handleOptions, json } from "../_shared/http.ts";
import { requireUser } from "../_shared/supabase.ts";

const OPENAI_API_KEY = Deno.env.get("OPENAI_API_KEY");
const OPENAI_MODEL = Deno.env.get("OPENAI_MODEL") ?? "gpt-5.2";

type RequestBody = {
  channel_id?: string;
  stop_names?: string[];
};

type OpenAiAlias = {
  original?: string;
  estimated?: string;
};

const prompt = `You are interpreting abbreviated station names from a rail replacement service display.

Each entry follows this structure:

* A 4-letter station code (derived from the real station name, often by removing vowels or compressing syllables)
* A 2-letter stop code (ignore this; it is ambiguous and not needed)
* An optional "T" indicating terminus (ignore for naming purposes)

Your task is to infer the full station names from the 4-letter codes.

Guidelines:

* Treat the 4-letter code as a compressed version of a real station name (e.g. consonant-heavy, missing vowels, or merged syllables)
* Use pattern recognition rather than strict decoding
* Prefer real-world plausibility over perfect letter matching
* Assume all stations are on the same rail corridor or geographically connected route
* Use the sequence of stops to inform your guesses (adjacent stations should make sense geographically)
* If a code is slightly irregular, prioritise what fits the route best over what matches the letters exactly

Output:
Return JSON with shape {"aliases":[{"original":"<code>","estimated":"<name>"}]}.
Return only valid JSON with no markdown or explanation.`;

function extractCode(rawStopName: string): string {
  const firstToken = rawStopName.trim().split(/\s+/)[0] ?? "";
  const normalized = firstToken.toUpperCase().replace(/[^A-Z0-9]/g, "");
  return normalized.slice(0, 8);
}

Deno.serve(async (req) => {
  const preflight = handleOptions(req);
  if (preflight) return preflight;

  if (req.method !== "POST") return fail("Method not allowed", 405);
  if (!OPENAI_API_KEY) return fail("OPENAI_API_KEY is not configured", 500);

  const { client, user, error: userError } = await requireUser(req);
  if (!user) return fail(userError ?? "Unauthorized", 401);

  let body: RequestBody;
  try {
    body = await req.json();
  } catch {
    return fail("Invalid JSON body");
  }

  const channelId = (body.channel_id ?? "").trim();
  const stopNames = Array.isArray(body.stop_names) ? body.stop_names : [];
  if (!channelId) return fail("channel_id is required");
  if (stopNames.length == 0) return json({ aliases: [] });
  if (stopNames.length > 1000) return fail("stop_names is too large");

  const { data: channel, error: channelError } = await client
    .from("channels")
    .select("id, type")
    .eq("id", channelId)
    .eq("type", "operations")
    .maybeSingle();
  if (channelError) return fail(channelError.message, 400);
  if (!channel) return fail("forbidden", 403);

  const rawByCode = new Map<string, Set<string>>();
  for (const stopName of stopNames) {
    const raw = `${stopName ?? ""}`.trim();
    if (!raw) continue;
    const code = extractCode(raw);
    if (!code) continue;
    const existing = rawByCode.get(code) ?? new Set<string>();
    existing.add(raw);
    rawByCode.set(code, existing);
  }

  const codes = [...rawByCode.keys()];
  if (codes.length == 0) return json({ aliases: [] });

  const openAiResponse = await fetch("https://api.openai.com/v1/chat/completions", {
    method: "POST",
    headers: {
      Authorization: `Bearer ${OPENAI_API_KEY}`,
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      model: OPENAI_MODEL,
      temperature: 0,
      response_format: { type: "json_object" },
      messages: [
        { role: "system", content: prompt },
        {
          role: "user",
          content: JSON.stringify({
            channel_id: channelId,
            codes,
          }),
        },
      ],
    }),
  });

  if (!openAiResponse.ok) {
    const details = await openAiResponse.text();
    return fail(`OpenAI request failed: ${details}`, 502);
  }

  const completion = await openAiResponse.json();
  const content = completion?.choices?.[0]?.message?.content;
  if (typeof content !== "string" || content.trim().length == 0) {
    return fail("OpenAI returned no content", 502);
  }

  let parsed: { aliases?: OpenAiAlias[] };
  try {
    parsed = JSON.parse(content);
  } catch {
    return fail("OpenAI returned invalid JSON", 502);
  }

  const codeToEstimated = new Map<string, string>();
  for (const row of parsed.aliases ?? []) {
    const original = `${row.original ?? ""}`.trim().toUpperCase();
    const estimated = `${row.estimated ?? ""}`.trim();
    if (!original || !estimated) continue;
    codeToEstimated.set(original, estimated);
  }

  const aliases = [];
  for (const [code, rawStops] of rawByCode.entries()) {
    const estimated = codeToEstimated.get(code);
    if (!estimated) continue;
    for (const raw of rawStops) {
      aliases.push({
        raw_stop_name: raw,
        alias_stop_name: estimated,
        source: "ai",
      });
    }
  }

  return json({ aliases });
});