## Business Pain Points
Business Pain Points
JD-to-resume matching on job platforms looks like a text similarity problem on the surface, but it's actually a high-frequency, high-cost, time-sensitive operational task. Take a mid-sized platform with 800 daily active employers as an example:
- About 1,200 new JDs are posted daily, with roughly 100,000 resumes submitted;
- Manual screening takes an average of 90 seconds per resume — 100,000 resumes means 2,500 labor hours, which translates to over 30,000 RMB per day based on first-tier city customer service costs;
- Employers' core demand is "getting feedback after applying," but the larger the resume pool, the longer the response delay. Every day of slow matching noticeably increases candidate churn.
Small and mid-sized platforms don't have the budget to train their own embedding models, nor the manpower to maintain integrations with multiple LLMs. The pain points boil down to three:
- High volume: Each resume needs bidirectional matching against multiple candidate JDs, with millions of vector comparisons per day;
- Heterogeneity: Resumes come as a mix of PDF/Word/images, while JDs are semi-structured text — the quality of field extraction directly caps the matching quality;
- Volatility: LLM pricing and capabilities change monthly. Locking into a single vendor makes both cost and availability uncontrollable.
Architecture Design
We broke the solution into four layers, assembled entirely from managed services + an API gateway. A team of two shipped it in three weeks:
┌─────────────────────────────────────────────┐
│ 接入层:JD 发布 / 简历上传 Webhook │
├─────────────────────────────────────────────┤
│ 解析层:文档转文本 → 结构化抽取 │
│ (技能、年限、行业、薪资期望、地点) │
├─────────────────────────────────────────────┤
│ 匹配层:Embedding 向量化 + 向量库检索召回 │
│ → LLM 精排打分(0-100 + 缺失项说明) │
├─────────────────────────────────────────────┤
│ 统一 AI API 网关:多模型路由、缓存、限流、 │
│ 降级、用量统计 │
└─────────────────────────────────────────────┘There are three key design decisions:
Separate retrieval from re-ranking. Retrieval uses a cheap embedding model for vector search — each JD only sends the Top 50 resumes to re-ranking. Re-ranking uses an LLM to output structured scores and reasons. This way, out of millions of calls, only one in a thousand actually hits the LLM, making token costs two orders of magnitude lower than a "dump everything into the LLM" approach.
Extraction results carry confidence scores. The most error-prone parts of resume parsing are years of experience and skill normalization ("3 years of Java" vs "three years of JAVA development experience"). The extraction layer outputs a confidence score for each field; anything below the threshold automatically degrades to "please confirm with the employer," avoiding complaints caused by incorrect matches.
All model calls go through a unified gateway. This is the key to low maintenance costs down the road, so let's expand on it.
Key Implementation Steps
The three-week schedule was roughly:
Week 1: Data pipeline and parsing
- Integrate a resume parsing service that outputs a unified JSON Schema (role, skills, years, industry, location, expect_salary);
- On the JD side, use a prompt to extract the same Schema. Once the fields on both sides are aligned, matching has a common coordinate system.
Week 2: Retrieval and re-ranking
- Generate embeddings for resumes and JDs respectively, and write them to a vector database;
- Complete re-ranking with a single prompt — the key is requiring the model to output JSON rather than free text.
Example re-ranking prompt call (pseudocode):
score = gateway.chat(
model="gpt-4o-mini", # 网关注册名,切换真实模型不改代码
response_format="json",
messages=[{
"role": "system",
"content": "你是招聘匹配专家。根据JD和简历JSON,"
"输出 {score: 0-100, reasons: [], gaps: []}"
}, {
"role": "user",
"content": f"JD: {jd_json}\n简历: {resume_json}"
}],
cache_key=hash(jd_json, resume_json), # 网关层缓存,重复投递不重复计费
fallback=["qwen-plus", "gemini-flash"] # 主模型限流时自动降级
)Week 3: Feedback loop and launch
- Feed back employers' "viewed / ignored / scheduled interview" actions as score calibration signals;
- Run weekly prompt regression tests against the top 20% of matches by interview rate, to prevent score drift after model updates.
Efficiency Comparison
Before/after comparison three months post-launch (based on internal platform operational data):
| Metric | Before (manual screening) | After (AI pipeline) |
|---|---|---|
| Screening time per resume | ~90 seconds | 0.8 seconds (including parsing) |
| Daily screening cost | ~30,000 RMB in labor | ~900 RMB in API fees |
| Submission to first feedback | 36 hours on average | 4 minutes on average |
| Employer weekly retention | — | Noticeable improvement, mainly due to earlier interview scheduling |
Of the API costs, about 30% of re-ranking calls hit the gateway cache (candidates repeatedly applying to similar positions is very common) — this alone saves nearly 10,000 RMB per month.
Why a Unified AI API Gateway Reduces Maintenance Costs
In this solution, we integrated 1 embedding model and 3 chat models (1 primary + 2 fallbacks). If we had connected to each vendor's SDK directly, every model switch would require changes to four types of code: authentication, request body, response parsing, and error retry — at least half a day by past experience. And vendor details scattered throughout business code would double testing costs.
The value of a unified gateway:
- Integrate once, switch models freely. Business code targets a single endpoint with a unified request format; switching models means changing one registered name, compressing model trials from half a day to a few minutes;
- Fallback and rate limiting aren't baked into business logic. When the primary model hits a 429 or times out, it automatically switches to a backup — no manual on-call during hiring peak seasons;
- Caching and usage analytics come built-in. Cache hits for repeated matches and token consumption per business are reported directly at the gateway layer — cost attribution requires no extra instrumentation;
- What small teams lack most is maintenance bandwidth. The gateway consolidates all this "non-business but must-be-done-right" work into a single dependency, enabling a two-person team to devote all their energy to extraction quality and score calibration.
Summary
The moat of matching-type AI applications isn't the model itself, but Schema design, the retrieval/re-ranking split, and the feedback loop. In this pipeline, the code you actually need to write is under two thousand lines — everything else is handled by managed services and the gateway. If your application also involves multi-model calls, high request frequency, and cost sensitivity, I recommend starting with the unified gateway layer. You can get it running just by registering an account: https://api.thistoken.ai/register
---
Ready to try it yourself? Sign up at https://api.thistoken.ai/register to get your API key and start building.
Bạn muốn thử Token.AI?
Tạo API Key cấp dự án, bật kênh trong bảng điều khiển và định cấu hình định tuyến, ngân sách và nhật ký kiểm tra.
注册 ThisToken.AI 并获取 API Key