AAIDC 2025 — AI Hiring Assistant RAG (readytensor_rag_starter)

A production-ready Retrieval-Augmented Generation (RAG) starter built for evaluating job candidates via resumes and job descriptions. Integrates embeddings, FAISS, and LLMs for semantic search and candidate ranking.

🌍 Problem Statement

Hiring processes often involve reviewing hundreds of resumes against complex job descriptions.
This manual process is time-consuming, error-prone, and subjective, leading to:

• Missed qualified candidates.
• Bias introduced by inconsistent human judgment.
• High recruiter workload and increased hiring costs.

Stakeholders: HR managers, recruiters, job candidates, and organizations scaling their workforce.

Impact: On average, recruiters spend 6–7 seconds scanning a resume. Scaling this to thousands of applicants results in inefficiency and potential loss of talent. Existing Applicant Tracking Systems (ATS) rely heavily on keyword matching, which fails to capture semantic meaning and contextual fit.

This project introduces an AI-powered assistant that leverages RAG pipelines to match resumes and job descriptions semantically, providing ranked candidate recommendations.

---

📚 Related Work & Gaps

• Existing HR Tech Tools (e.g., ATS, LinkedIn Recruiter) use keyword or rule-based filters.
• Gaps:
  • Limited semantic understanding of resumes/job descriptions.
  • Poor adaptability to domain-specific roles (e.g., AR/VR specialists, MetaHuman experts).
  • Lack of transparent pipelines for evaluation and reproducibility.

Our solution addresses these gaps by combining FAISS vector stores, sentence-transformer embeddings, and LLM reasoning to provide semantic candidate ranking.

---

Features

• FastAPI API with /ingest, /ingest_resumes, and /ask routes
• Pluggable embeddings (local sentence-transformers or Gemini embeddings)
• FAISS vector store persisted to disk
• Transparent RAG pipeline (chunk → embed → index → retrieve → generate)
• Configurable via .env
• Bulk ingestion of resumes and job descriptions
• Swagger UI /docs and ReDoc /redoc support

---

Quickstart

1) Clone & create environment

git clone <your-repo-url>
cd readytensor_rag_starter
python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

2) Configure environment

Copy .env.example to .env and set values:

cp .env.example .env

Key variables:

• EMBEDDING_PROVIDER — "local" or "gemini"
• LLM_PROVIDER — "gemini" (or local/testing)
• EMBEDDING_MODEL — e.g., sentence-transformers/all-MiniLM-L6-v2
• MODEL_NAME — e.g., gemini-2.5-flash-lite
• GEMINI_API_KEY — if using Gemini
• CHUNK_SIZE, CHUNK_OVERLAP, TOP_K, MAX_TOKENS, TEMPERATURE

3) Ingest resumes and job descriptions

Bulk ingestion:

python -m rag.ingest --source data/resumes_and_jd

Single file ingestion:

curl -X POST "http://localhost:8000/ingest" \
-H "Content-Type: application/json" \
-d '{"file_path":"data/resumes_and_jd/jake.pdf"}'

This will build/persist the FAISS index and metadata for semantic search.

4) Run the API

uvicorn app.main:app --reload

Swagger UI: http://localhost:8000/docs
ReDoc: http://localhost:8000/redoc

5) Ask a question

curl -X POST "http://localhost:8000/ask" \
-H "Content-Type: application/json" \
-d '{"question":"Rank candidates for the MetaHuman Specialist role","top_k":4}'

---

Project Structure

readytensor_rag_starter/
├── app/
│   └── main.py                     # FastAPI app
├── rag/
│   ├── ingest.py                   # Resume/job ingestion
│   ├── pipeline.py                 # RAG pipeline logic
│   ├── prompt.py                   # Prompts for candidate evaluation
│   └── utils.py                    # Helpers (PDF parsing, embeddings, vector store)
├── data/
│   └── resumes_and_jd/             # PDFs, job descriptions, candidates.csv
├── storage/                        # generated FAISS index + metadata (runtime)
├── .env_example
├── .gitignore
├── requirements.txt
└── README.md

---

API Endpoints

---

Usage Tips

• Place all resumes and job descriptions in data/resumes_and_jd/
• Run python -m rag.ingest --source data/resumes_and_jd before querying
• Start the API server with uvicorn app.main:app --reload
• Query /ask to evaluate or rank candidates

---

Limitations & Trade-offs

• No built-in bias mitigation — relies on embeddings/LLM neutrality.
• Performance depends on model choice (sentence-transformers vs Gemini).
• Current implementation is API-only (no recruiter dashboard).
• Limited testing on large-scale datasets (scalability still needs benchmarks).

---

Performance & Scalability

*Tested locally with ~20 candidate resumes and 1 job description.
*Indexing: ~2–3s per document on CPU.
*Query latency: ~1–2s with local embeddings.
*Resource needs: ~2GB RAM minimum; GPU recommended for large datasets.
*Scalability: FAISS supports millions of vectors, but not yet benchmarked for >10k resumes in this repo.

---

Maintenance & Support

*Current Version: v0.1.0 (experimental).
*Updates: Manual, planned as new features are added.
*Support: Open an issue on GitHub Issues.

---

Visual Demonstration

*
*

Contact

*Maintainer: Razwi M K
*GitHub: @razwi1
*Email: razwimkofficial@gmail.com

---

Requirements

• fastapi
• uvicorn[standard]
• python-multipart
• pydantic
• openai
• langchain
• tiktoken
• pandas
• numpy
• python-docx
• PyPDF2
• sentence-transformers
• faiss-cpu

---

Make Targets

make install       # pip install -r requirements.txt
make ingest        # bulk ingest resumes/job descriptions
make dev           # start uvicorn server
make test          # run tests
make format        # black + isort formatting

---

Security & Reproducibility

• Keep secrets in .env (never commit). .env_example provided.
• FAISS indexes are deterministic given the same inputs.
• Pin dependency versions in requirements.txt for reproducibility.

---

License

MIT License