Skip to main content
Generate vector embeddings for semantic search, clustering, and similarity tasks.

​
Job Types

From work.hpp:15-19, nrvna-ai supports three job types: 
enum class JobType : uint8_t {
    Text = 0,
    Embed = 1,
    Vision = 2
};
Use JobType::Embed to generate embeddings instead of text completions.

​
Basic Usage

Submit an embedding job: 
wrk ./workspace "Search query text" --type embed
The result is a vector (array of floats) instead of generated text.

​
API Reference

From work.hpp:46: 
SubmitResult submit(const std::string& prompt, JobType type = JobType::Text);
Specify JobType::Embed to generate embeddings.

​
Batch Embeddings

Generate embeddings for multiple texts: 
# Embed a corpus
while IFS= read -r line; do
  wrk ./workspace "$line" --type embed >> jobs.txt
done < corpus.txt

# Collect vectors
for job in $(cat jobs.txt); do
  flw ./workspace $job >> embeddings.json
done

​
Use Cases

1

Semantic search

Embed documents and queries, then compute cosine similarity:
# Embed documents
for doc in docs/*.txt; do
  wrk ./workspace "$(cat $doc)" --type embed >> doc-embeddings.txt
done

# Embed query
query_emb=$(wrk ./workspace "search query" --type embed | xargs flw ./workspace)

# Compare with similarity function
# (compute cosine similarity in your app)
2

Clustering

Group similar texts by embedding distance:
# Embed all items
for item in items/*.txt; do
  wrk ./workspace "$(cat $item)" --type embed
done

# Cluster vectors with k-means or DBSCAN
# (use Python/NumPy for clustering)
3

Duplicate detection

Find near-duplicate content:
# Embed all candidates
for candidate in candidates/*.txt; do
  id=$(basename "$candidate" .txt)
  emb=$(wrk ./workspace "$(cat $candidate)" --type embed | xargs flw ./workspace)
  echo "$id|$emb" >> embeddings.txt
done

# Find pairs with high similarity
# (threshold cosine similarity > 0.95)

​
Model Selection

Use embedding-specialized models for best results: 
# Dedicated embedding workspace
nrvnad nomic-embed-text.gguf ./ws-embed &

# Submit to embedding workspace
wrk ./ws-embed "text to embed" --type embed
Embedding models are typically smaller and faster than generative models.

​
Configuration

Embedding jobs don’t generate tokens, so adjust settings: 
# No need for large predict size
export NRVNA_PREDICT=1

# Context size = max input length
export NRVNA_MAX_CTX=2048

# Max workers for throughput
export NRVNA_WORKERS=8

nrvnad embed-model.gguf ./workspace

​
Output Format

Embedding results are vectors (arrays of floats). Parse them in your application: 
import json
import subprocess

# Submit and wait
job_id = subprocess.check_output(['wrk', './workspace', 'text', '--type', 'embed']).strip()
result = subprocess.check_output(['flw', './workspace', job_id])

# Parse vector
vector = json.loads(result)
print(f"Embedding dimension: {len(vector)}")

​
Tips

  • Embedding vs. Text — use --type embed for vectors, default for completions
  • Model choice — embedding models outperform generative models on similarity tasks
  • Normalization — some models return normalized vectors, others don’t
  • Dimensions — typical sizes are 384, 768, 1024, or 1536
  • Batch processing — embeddings are fast; process thousands per minute

Build docs developers (and LLMs) love

Get started for freeTalk to us