Skip to main content

​
Synthetic Faces

A synthetic face is defined by a boolean formula over existing faces. Its psychological primitives are assembled on the fly — no separate compilation, no extra training. The formula is evaluated at inference time by taking the union, intersection, or difference of the contributing faces’ compiled knowledge.

​
Formula syntax

OperatorMeaning
a | bUnion — everything from a and everything from b
a & bIntersection — only what a and b share
a - bDifference — everything from a that is not in b
a ^ bSymmetric difference — everything unique to either a or b
Parentheses are supported for grouping: 
(alice | bob) & mentor
alice | (expert - academic)

​
Create a synthetic face

Pass a formula instead of uploading documents: 
curl -X POST https://api.faces.sh/v1/faces \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "username": "alice-and-bob",
    "formula": "alice | bob"
  }'
The formula references faces by their username. All referenced faces must:
  • Be owned by you
  • Be concrete (non-synthetic) faces
Nesting synthetic faces — referencing a face that itself has a formula — is not supported and returns 422.

​
Update the formula

curl -X PATCH https://api.faces.sh/v1/faces/FACE_ID \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"formula": "alice & mentor"}'
Changes take effect immediately on the next chat request.

​
Chat with a synthetic face

Exactly the same as chatting with any face: 
curl -X POST https://api.faces.sh/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "alice-and-bob",
    "messages": [{"role": "user", "content": "How do you approach creative problems?"}]
  }'

​
Validation rules

  • All referenced usernames must be owned by you at create/update time.
  • Referenced faces must be concrete (not synthetic themselves).
  • The formula must be a valid boolean expression using |, &, -, ^, parentheses, and valid usernames.
Invalid formulas return 422 Unprocessable Entity with a description of the problem.

​
How it works

When a chat request arrives for a synthetic face, the profiler service:
  1. Evaluates the formula, reading the compiled knowledge graphs for each referenced face.
  2. Applies the boolean operations to produce a merged knowledge set.
  3. Retrieves and ranks the most relevant primitives from the merged set.
  4. Injects them into the system prompt as if they were compiled for a single face.
The result is a coherent persona that blends the source faces according to the formula — without any training or separate compilation step.

​
Use cases

  • Blended expertise: designer | engineer — draw on both knowledge domains in a single persona.
  • Constrained character: alice & mentor — only the facets of Alice that overlap with a mentor archetype.
  • Differenced persona: senior-dev - junior-dev — the characteristics that distinguish a senior developer from a junior one.
  • A/B testing: Create alice | bob and alice & bob to compare unioned vs. intersected personas.