Embeddings

Convert text into numerical vectors for semantic search, similarity comparison, and clustering

Endpoint

POST /llm/v1/embeddings

cURL

curl -X POST https://api.case.dev/llm/v1/embeddings \
  -H "Authorization: Bearer sk_case_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/text-embedding-3-small",
    "input": "Plaintiff alleges negligence in post-operative care"
  }'

Response

{
  "object": "list",
  "data": [
    {
      "object": "embedding",
      "index": 0,
      "embedding": [-0.016, 0.022, -0.011, ...]
    }
  ],
  "model": "openai/text-embedding-3-small",
  "usage": {
    "prompt_tokens": 12,
    "total_tokens": 12
  }
}

Parameters

Parameter	Type	Required	Description
model	string	Yes	Embedding model ID
input	string or array	Yes	Text(s) to embed

Embedding models

Model	Dimensions	Best for	$/1K tokens
openai/text-embedding-3-small	1536	General purpose	$0.00002
openai/text-embedding-3-large	3072	Higher quality	$0.00013
voyage/voyage-law-2	1024	Legal documents	$0.00012
voyage/voyage-3.5	1536	General purpose	$0.00006

For legal documents, use voyage/voyage-law-2. It’s specifically trained on legal text.

Batch embeddings

Embed multiple texts in one request:

CLI

casedev llm:v1 create-embedding \
  --model openai/text-embedding-3-small \
  --input "Plaintiff alleges negligence in post-operative care"

Use cases

Semantic search

CLI

# 1. Embed your query
casedev llm:v1 create-embedding \
  --model voyage/voyage-law-2 \
  --input "negligence in surgical procedure"

# 2. Compare to document embeddings (cosine similarity)
# 3. Return most similar documents

Document similarity

CLI

# Compare two documents (embed each separately, then compute similarity)
casedev llm:v1 create-embedding \
  --model voyage/voyage-law-2 \
  --input "Plaintiff expert testimony on standard of care"

casedev llm:v1 create-embedding \
  --model voyage/voyage-law-2 \
  --input "Defense expert rebuttal on treatment protocols"

Tip: Use the same model for indexing and querying. Mixing models produces poor results.