From e84871926bd4a9f33851975e1a520115b93ccc3d Mon Sep 17 00:00:00 2001
From: JobsBoard Deployer <admin@mithraic.cloud>
Date: Sat, 5 Sep 2026 14:02:58 -0400
Subject: [PATCH] Enable high-precision OpenRouter LLM extraction for resume
 parsing to prevent scrambled titles and organization names

---
 web/src/app/api/resume/parse/route.ts | 86 +++++++++++++++++++++++----
 1 file changed, 73 insertions(+), 13 deletions(-)

diff --git a/web/src/app/api/resume/parse/route.ts b/web/src/app/api/resume/parse/route.ts
index 7658787..9562306 100644
--- a/web/src/app/api/resume/parse/route.ts
+++ b/web/src/app/api/resume/parse/route.ts
@@ -340,7 +340,10 @@ function parseResumeFull(rawText: string) {
 
 async function parseResumeWithLLM(rawText: string) {
   const apiKey = process.env.OPENROUTER_API_KEY || process.env.OPENAI_API_KEY;
-  if (!apiKey) return null;
+  if (!apiKey) {
+    console.log("[Resume Parser] No OPENROUTER_API_KEY or OPENAI_API_KEY configured; using heuristic parser.");
+    return null;
+  }
 
   const isRouter = !!process.env.OPENROUTER_API_KEY;
   const endpoint = isRouter
@@ -348,38 +351,94 @@ async function parseResumeWithLLM(rawText: string) {
     : `${process.env.OPENAI_BASE_URL || "https://api.openai.com/v1"}/chat/completions`;
   const model = process.env.OPENROUTER_MODEL || process.env.OPENAI_MODEL || (isRouter ? "google/gemini-2.5-flash" : "gpt-4o-mini");
 
+  console.log(`[Resume Parser] Invoking LLM parser via ${endpoint} using model: ${model}`);
+
+  const systemPrompt = `You are a high-precision ATS resume parser and information extraction engine.
+CRITICAL EXTRACTION GUIDELINES:
+1. Parse raw, messy, or multi-column resume text into clean, structured JSON.
+2. Accurately untangle company names and position titles:
+   - Identify the real organization/employer (e.g. "Pokemoto", "Christmas Wish CT", "Personal Homelab").
+   - NEVER output generic placeholders like "Organization" or "Company" if the actual organization name is stated anywhere nearby.
+   - Do NOT place dates or location text into the company name field.
+3. Identify the candidate's actual name, email, phone number, and location accurately.
+4. Extract every distinct job experience entry with its authentic position, company, and bullet points (highlights).
+5. Extract explicit skills into the skills array.
+6. Extract schools/universities and degrees into the education array.
+7. Return strictly a JSON object with this schema:
+{
+  "profile": {
+    "name": string,
+    "headline": string,
+    "email": string,
+    "phone": string,
+    "location": string,
+    "summary": string
+  },
+  "work": [
+    {
+      "position": string,
+      "company": string,
+      "summary": string,
+      "highlights": string[]
+    }
+  ],
+  "skills": string[],
+  "education": [
+    {
+      "degree": string,
+      "institution": string
+    }
+  ]
+}`;
+
   try {
+    const controller = new AbortController();
+    const timeout = setTimeout(() => controller.abort(), 15000);
+
     const res = await fetch(endpoint, {
       method: "POST",
       headers: {
         Authorization: `Bearer ${apiKey}`,
         "Content-Type": "application/json",
+        ...(isRouter ? { "HTTP-Referer": "https://directwire.io", "X-Title": "DirectWire Resume Parser" } : {}),
       },
+      signal: controller.signal,
       body: JSON.stringify({
         model,
+        temperature: 0.1,
         messages: [
-          {
-            role: "system",
-            content:
-              "You are a professional ATS resume parser. Extract structured JSON strictly matching: { profile: { name: string, headline: string, email: string, phone: string, location: string, summary: string }, work: [{ position: string, company: string, summary: string, highlights: string[] }], skills: string[], education: [{ degree: string, institution: string }] }",
-          },
+          { role: "system", content: systemPrompt },
           {
             role: "user",
-            content: `Extract structured JSON resume data:\n\n${rawText.slice(0, 10000)}`,
+            content: `Extract structured JSON resume from this text:\n\n${rawText.slice(0, 12000)}`,
           },
         ],
         response_format: { type: "json_object" },
       }),
     });
 
-    if (!res.ok) return null;
+    clearTimeout(timeout);
+
+    if (!res.ok) {
+      console.warn("[Resume Parser] LLM API responded with error:", res.status, await res.text());
+      return null;
+    }
+
     const json = await res.json();
-    const content = json.choices?.[0]?.message?.content;
+    let content = json.choices?.[0]?.message?.content;
     if (!content) return null;
 
-    return JSON.parse(content);
-  } catch (e) {
-    console.error("LLM parse fallback triggered:", e);
+    // Clean markdown code fence wrappers if present
+    content = content.replace(/^```(?:json)?\s*/i, "").replace(/\s*```$/i, "").trim();
+
+    const parsed = JSON.parse(content);
+    if (parsed && (parsed.profile || Array.isArray(parsed.work))) {
+      console.log("[Resume Parser] Successfully extracted structured resume with LLM.");
+      return parsed;
+    }
+    return null;
+  } catch (e: any) {
+    console.error("[Resume Parser] LLM parse call failed or timed out:", e.message);
     return null;
   }
 }
@@ -469,7 +528,8 @@ export async function POST(req: Request) {
 
     // Try LLM parsing first if API key configured, otherwise use high-precision local parsing
     let parsedData = await parseResumeWithLLM(rawText);
-    if (!parsedData || !parsedData.profile || !parsedData.profile.name) {
+    if (!parsedData || !parsedData.profile || (!parsedData.profile.name && (!Array.isArray(parsedData.work) || parsedData.work.length === 0))) {
+      console.log("[Resume Parser] LLM returned empty or incomplete data; applying local heuristic extraction.");
       parsedData = parseResumeFull(rawText);
     }