epithre-embed

Multimodal embedding model. Indonesian-optimized, with text and image vectors in a shared 4000-dim space.

Capabilities

When to use

• RAG retrieval (text + image)
• Semantic search
• Clustering / deduplication
• Classification via centroid distance
• Cross-modal search (find images by text query, or vice versa)

Key feature: cross-modal in one space

Text and image embeddings live in the same 4000-dim space. Cosine similarity between any text vector and any image vector is meaningful, so you can:

• Search an image catalog by text description.
• Find text passages matching an uploaded photo.
• Maintain a unified pgvector index for mixed-modality corpora (one table, one index, one similarity function).

Output details

• Vectors are L2-normalized (cosine sim = dot product)
• 4000 native dims, Matryoshka-truncatable: dimensions=1024 returns the first 1024 dims re-normalized (lossless prefix)
• halfvec-compatible for pgvector storage

Pricing

• Text: Rp1,500 / 1M input tokens
• Image: Rp25 / image (flat)
• Batch: 0.5x both rates
• Storage: free (1 GB quota for uploaded knowledge files)

Performance

• Throughput: ~25-40 req/s sustained on a single backend GPU.
• Latency: ~0.1-0.5s for a batch of 64 short texts; ~0.5-1s for image input.
• Quality on Indonesian retrieval: ~92% recall@10 on our 25-question gold set.

Limits

• Max 64 text items per request
• Max 8 image items per request
• Max 25 MB request body
• Each text item auto-truncated to 10K chars (matched against the 4096-token context window)

Caveats

• instruction field is text-only. Mixed text + image input rejects instruction.
• Image embeddings are MRL-truncated to 4000 + L2-renormalized in the gateway to share the same space as text vectors. Effectively identical to native 4000-dim output.
• Truncation is char-based, not token-based. For dense legal/code text (2-3 chars/token), the 10K-char limit corresponds to ~3000-3500 tokens.

See also

• Embeddings reference
• Multimodal guide
• Cookbook: RAG, Cross-modal RAG