jobs-rag

A job recommendation engine that ingests RSS feeds, extracts structured data with an LLM, embeds listings with a text embedding model, stores them in Gel (EdgeDB), and surfaces results via vector similarity or LLM-powered reasoning search.

Architecture

RSS Feed
   ↓
parse_rss_feed()       →  JobEntry (title, link, summary, published)
   ↓
label()                →  Opportunity (structured fields via llama3.2)
   ↓
embed()                →  vector[768] (nomic-embed-text)
   ↓
Gel (EdgeDB)           →  stores Opportunity + embedding
   ↓
recommend / search     →  cosine similarity  |  LLM reasoning

Requirements

• Python 3.11+
• Poetry
• Gel CLI
• Ollama server with llama3.2 and nomic-embed-text pulled

Setup

1. Install dependencies

poetry install

2. Configure Ollama host

The Ollama server is configured in two places:

File	Variable	Purpose
ingestion/labeler.py	_client host	LLM labeling (llama3.2)
ingestion/embedder.py	_client host	Embeddings (nomic-embed-text)
reasoned-recommend.py	_llm host	LLM reasoning search

Update the host URLs to point at your Ollama instance.

3. Initialize the database

# Initialize a local Gel instance for this project
gel project init

# Apply migrations (creates the JobOpportunity type)
gel migrate

4. Ingest jobs

Fetches RSS feeds, labels each listing with the LLM, embeds it, and upserts into Gel.

poetry run python -m ingestion

Output per entry:

20:49:21 [INFO] ingestion.main: [1/25] Labeling: Finance Expert - Portfolio Management
20:49:26 [INFO] ingestion.main: [1/25] Embedding: Finance Expert @ xAI
20:49:27 [INFO] ingestion.main: [1/25] Saving to DB
20:49:27 [INFO] ingestion.main: [1/25] Done: Finance Expert @ xAI

Usage

Vector similarity search (fast)

Embeds the query and returns the closest matches by cosine similarity.

poetry run python recommend.py "remote senior finance role"

Reasoning search (slower, with explanations)

Retrieves the top N candidates by vector similarity, then passes them to the LLM to reason about fit and re-rank with explanations.

poetry run python reasoned-recommend.py "200k on-site software engineering job"

Output includes a fit score and reasoning per result:

#1  Finance Manager @ Acme Corp
    Location:  New York (on-site)
    Seniority: senior  |  Type: full-time
    Skills:    Excel, SQL, Financial Modeling
    Fit score: 0.87
    Reasoning: Strong match — senior finance role in a major city...

Project structure

jobs-rag/
├── dbschema/
│   ├── default.gel          # Gel schema
│   └── migrations/          # Applied migrations
├── ingestion/
│   ├── __main__.py          # Entry point: python -m ingestion
│   ├── main.py              # Pipeline orchestration
│   ├── labeler.py           # LLM structured extraction (llama3.2)
│   ├── embedder.py          # Text embeddings (nomic-embed-text)
│   ├── store.py             # Gel upsert
│   └── sources/
│       └── rss.py           # RSS parsing
├── recommend.py             # Vector similarity search
└── reasoned-recommend.py    # LLM reasoning search

Adding more RSS feeds

Edit RSS_FEEDS in ingestion/main.py:

RSS_FEEDS = [
    "https://weworkremotely.com/categories/remote-management-and-finance-jobs.rss",
    "https://example.com/jobs.rss",
]

Re-run poetry run python -m ingestion — existing entries are upserted by link, so no duplicates.