cvchecker: a Privacy-Preserving RAG CV Matcher

Overview

A lightweight, privacy-focused RAG pipeline wrapped in a FastAPI REST API. It evaluates CVs against job postings using:

• Local Sanitization: CPU-optimized local models (via Ollama) for privacy.
• Efficient Retrieval: FAISS for in-memory, lightning-fast semantic search.
• Frontier Reasoning: Google Gemini (Free Tier) for high-level analysis and scoring.

Prerequisites

A quick note about the prerequisites...

• Python 3.13 was used for the development of this project, so no guarantees are given that it will work with any other versions of Python (although it almost certainly will).
• Some parts of the pipeline depend on ollama running on the host where the API is deployed. Thus ollama shall be installed on the host.
• Other parts rely on Google's Gemini models, thus an appropriate GEMINI_API_KEY variable shall be set in the environment to a key from Google AI studio.
• To avoid hitting rate limits with HuggingFace (used for the embedding model and the test dataset), it's advisable to set the HF_TOKEN variable as well.

Build & run

Ensure the llama3.2:3b model is downloaded to the host running Ollama:

ollama pull llama3.2:3b

Tip

If you don't have Ollama running in the background, you can start it up in a separate terminal with ollama serve.

The uv package manager is used for this project, thus to pull all the dependencies, you'll first have to:

uv sync

Next, to get the system up and running locally in a developer environment:

fastapi dev

...or if you want to use it in a production environment (not recommended as of now):

fastapi run

By default, the API is exposed on http://localhost:8000/, but this can be configured to anything else (with the help of FastAPI) if needed.

Technical Stack

Component	Recommended Tool	Justification
Framework	FastAPI	High performance, minimal overhead, built-in async support.
Frontier LLM	Google Generative AI SDK	Access to Gemini Pro/Flash; generous free tier.
Local LLM	Ollama (Llama-3.2-3B)	Offloads processing to a local daemon; keeps Python env lean.
Vector DB	FAISS (cpu)	Industry standard for efficient, in-memory similarity search.
Embeddings	fastembed	Highly optimized for CPU (ONNX); significantly faster/lighter than sentence-transformers.
PDF Parsing	pypdf	Pure Python, lightweight, no external C-dependencies.
Web Scraping	httpx + selectolax	selectolax is much faster and uses less memory than BeautifulSoup.
Misc.	pydantic + dotenv	Needed to accomplish various tasks related to the main funtionality.

Detailed Pipeline Logic

Phase 1: Resume Processing (Local & Private)

1. Extraction: Parse the Resume PDF using pypdf into meaningful, self-contained text chunks (e.g., 500-1000 characters).
2. Sanitization (Local LLM via Ollama): Run each chunk through a local model (e.g., Llama-3.2) for:
   • PII Redaction: Identify and mask sensitive data (Names, Emails, Phone numbers). Use Regex for deterministic patterns first.
3. Indexing: Generate embeddings via fastembed and insert chunks into the Resume FAISS Index (in-memory).

Phase 2: Job Posting (JP) Processing

4. Acquisition: Fetch the JP from the provided URL, and scrape it using httpx and selectolax.
5. Indexing: Chunk the JP, generate embeddings, and insert into the JP FAISS Index (in-memory).
6. Requirement Analysis:
   • Perform a semantic search on the JP index to gather context.
   • Use Google Gemini to summarize the JP into a set of formalized requirements.
   • Output Structure: Each requirement MUST be formalized as a specific, answerable question about the candidate (e.g., "Does the candidate have 3+ years of experience with React?").

Phase 3: Matching & Evaluation (Frontier)

7. Requirement Verification: For each formalized question:
   • i) Perform a semantic search in the Resume FAISS Index to retrieve internal evidence chunks.
   • ii) Provide the External Evidence (from Phase 3.5) as additional grounding context.
   • iii) Use Google Gemini to check if the combined evidence confirms the requirement.
8. Final Evaluation:
   • Gemini aggregates evidences found in the CV into a fair, grounded summary.
   • Scoring: Compute a final match score (0-100). The score should be weighted higher if claims are verified by external links.
   • Output: Return a structured JSON response containing the summary, requirement list, verified links, and final score.

REST API Endpoints

• POST /check-cv: Upload PDF, sanitize, do the processing, and generate a report.