Vector search, neural search, and natural language queries one API, one embedding model, 4M+ live listings sourced directly from company career pages.
QUERY (NATURAL LANGUAGE)
"senior backend engineer building distributed payment systems with Go and Kafka, remote-friendly EU"
SOCRATES V2 EMBEDDING · 768 DIMENSIONS
[-0.0294, 0.0613, -0.0413, 0.0187, -0.0822, 0.0154, ...762 more]
Senior Payment Platform Engineer
Klarna · Stockholm · Remote EU
Staff Engineer, Money Movement
Wise · London · Hybrid
Senior Software Engineer, Payments Infrastructure
Adyen · Amsterdam · Hybrid
Pure similarity, hybrid with filters, or raw vector input — all backed by the same socrates v2 index.
Pass a natural language query, a job ID, or your own pre-computed vectors. Get the top-K most similar listings ranked by cosine similarity.
Send a natural language query under vector and structured filters under lexical — location, salary, experience, industry, posting date. Combined ranking, single round trip.
Pass raw 768-dim vectors directly via the vectors parameter. Skip re-encoding when you've cached embeddings in your own pipeline.
What the Semantic Search API actually powers in production.
Let users type "remote backend role at a Series B fintech with equity" instead of clicking through filters. The API matches role, seniority, company stage, geography, and compensation as a single query.
Pass a job ID, get the top-K most similar listings. ML engineer roles cluster with Applied Researcher; data scientist clusters with quant analyst. Powers detail-page rails and "you might also like" feeds.
Feed user intent — natural language goals, conversation context, scraped requirements — directly to the API. Match user descriptions to live listings without prompt engineering against keywords.
"ML engineer" should match Applied Researcher AND filter to remote, $150k+, senior, posted in the last 30 days. Neural Search runs both in one request with combined ranking.
Cache vectors once, query them many times. Match cached user-intent vectors against fresh job postings every hour without re-encoding. Powers batch matching, scheduled job alerts, and custom recommender stacks.
Page 1 of every search endpoint returns results without an API key. Prototype before you sign up.
Works with any HTTP client. Official Python SDK — pip install hirebase.
Every result returns a similarity score 0.0 - 1.0. Pass score to drop low-quality matches.
# Filtered semantic search in one request
curl -X POST https://api.hirebase.org/v2/jobs/neural-search \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_API_KEY" \
-d '{
"vector": {
"query": "senior backend engineer building
distributed systems with Go and Kafka"
},
"lexical": {
"location_types": ["Remote", "Hybrid"],
"experience": ["Senior"],
"industry": "Tech, Software & IT Services",
"days_ago": 30,
"limit": 10
}
}'# Returns ranked results with similarity scores
{
"jobs": [
{
"_id": "6814bw99fc2284gt4777f21a",
"job_title": "Senior Payment Platform Engineer",
"company_name": "Klarna",
"location_type": "Remote",
"experience_level": "Senior",
"vector_score": 0.92,
"date_posted": "2026-04-15"
}, ...
],
"total_count": 847,
"total_pages": 85
}Domain-tuned embeddings preserve role, seniority, function, and skill-level distinctions that general-purpose models flatten.
[ 01 ] Embedding model
Proprietary 768-dimensional model trained on tens of millions of job postings. Model name and version returned on every embedding response so you can pin to a specific release.
[ 02 ] Similarity scoring
Vector Search returns `score`; Neural Search returns `vector_score`. Threshold low-quality results with the optional `score` request parameter.
[ 03 ] Three input modes
Pass natural language (encoded server-side), the Mongo ObjectId of a reference job (uses the precomputed embedding), or your own raw 768-dimensional vector. Same endpoint, three workflows.
[ 04 ] Accuracy controls
The `accuracy` parameter (`low` / `medium` / `high`) trades exhaustiveness for response time. Default is `medium`.
[ 05 ] Result count
Use `top_k` to control candidate set size before pagination — useful for re-ranking with your own model. Pairs with `limit`, `page`, or `offset`.
[ 06 ] Free first page
Both Vector Search and Neural Search return page 1 without authentication. Prototype before signing up. Page 2+ requires a free key.
Richer prompts give the embedding model more dimensions to match against. A well-written query consistently outperforms structured filters alone.
Don't relegate context to filter parameters when the embedding model can match it from the query itself.
Stage, size, and industry character match through the company entity linkage on every job.
For agent and copilot workflows, give the model both the candidate's current state and their target.
Vector Search accepts lexical parameters but does not currently honor them. For 30-day windows, salary floors, industry, or location filtering to actually narrow results, use Neural Search with both vector and lexical blocks.
vs. building it yourself with general-purpose embeddings and a vector DB, or buying keyword-only job posting feeds.
Page 1 is free. Run a real semantic search in your terminal in 60 seconds.
