Documentation

Base URL

https://api.parallelfoundry.com/v1

Model

parallel-foundry/parallel

Key

header Authorization: Bearer pf-... (or X-API-Key); make one on Keys in the console

Price

$0.025 per million input tokens; output is free; no minimum per call (how tokens are counted)

curl https://api.parallelfoundry.com/v1/decide \
  -H "Authorization: Bearer $PARALLEL_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "My card was charged twice, please refund one payment.",
    "heads": ["intent=refund,card lost,balance", "urgent=bool", "frustration=score:1-5"]
  }'

{
  "intent": {"type": "choice", "answer": "refund",
             "probabilities": {"refund": 0.9942, "balance": 0.0037, "card lost": 0.0021}},
  "urgent": {"type": "bool", "answer": "no", "p_yes": 0.3758},
  "frustration": {"type": "score", "answer": "2", "expected": 2.263,
                  "probabilities": {"2": 0.3698, "1": 0.2707, "3": 0.2189, "4": 0.1072, "5": 0.0334}},
  "usage": {"input_tokens": 76, "output_tokens": 0,
            "input_tokens_detail": {"decision": 76, "media": 0, "writing": 0, "unread": 0}},
  "seconds": 0.021
}

One request: an input (a text, images, a recording, or a text with them) and any number of questions about it, all answered in one pass. The questions can be strings or objects.

{
  "text": "The food was great but we waited 40 minutes.",
  "heads": [
    {"name": "sentiment", "type": "choice", "question": "overall sentiment", "options": ["positive", "mixed", "negative"]},
    {"name": "urgent", "type": "bool", "question": "Is it urgent?"},
    {"name": "rating", "type": "score", "question": "rating", "options": ["1", "2", "3", "4", "5"]},
    {"name": "aspects", "type": "multi", "question": "aspects", "options": ["food", "service", "price", "wait"]}
  ],
  "reason": true
}

{
  "images": ["data:image/jpeg;base64,/9j/4AAQ..."],
  "image_tokens": 280,
  "heads": ["type=invoice,receipt,letter,form", "amount_over_300=bool"]
}

{
  "audio": ["data:audio/wav;base64,UklGR..."],
  "heads": ["intent=card issues,balance,pay bill,other"]
}

{
  "text": "Can you book a table for four tomorrow at 7:30 pm at Luma Bistro?",
  "tools": [
    {"name": "book_table", "description": "Book a restaurant table",
     "parameters": {"type": "object", "properties": {"restaurant": {"type": "string"},
                    "people": {"type": "integer"}, "time": {"type": "string"}}, "required": ["restaurant", "people"]}},
    {"name": "get_weather", "description": "Current weather in a city",
     "parameters": {"type": "object", "properties": {"city": {"type": "string"}}}}
  ],
  "tool_choice": "auto"
}

One entry per question, under its name, and the call's usage:

The OpenAI SDKs work with base_url="https://api.parallelfoundry.com/v1". Two kinds of call:

• Decisions: response_format json_schema whose fields are enums, booleans, integer ranges (2 to 11 values) or arrays of enums. Every field is decided in one pass; the content is that JSON, always valid, and decisions holds every option's %. A field's description is its question. Images (image_url, base64) and recordings (input_audio) work too.
• Tool calls: tools and tool_choice: the tool is picked with %, its arguments written; when no tool fits, the answer is empty with each tool's %.

Free-form answers (plain chat, written JSON, the answer after a tool result) are not offered: they get 400 chat_not_offered. stream: true works (with stream_options.include_usage). GET /v1/models lists the model; GET /v1/key shows your key's credit and limits.

from openai import OpenAI

client = OpenAI(base_url="https://api.parallelfoundry.com/v1", api_key="pf-...")
r = client.chat.completions.create(
    model="parallel-foundry/parallel",
    messages=[{"role": "user", "content": "I was charged twice, please refund one payment now!"}],
    response_format={"type": "json_schema", "json_schema": {"name": "ticket", "schema": {
        "type": "object", "properties": {
            "intent": {"type": "string", "enum": ["refund", "cancel order", "order status", "complaint"]},
            "urgent": {"type": "boolean"},
            "frustration": {"type": "integer", "minimum": 1, "maximum": 5}}}}})
print(r.choices[0].message.content)   # {"intent": "refund", "urgent": true, "frustration": 4}
print(r.model_dump()["decisions"])    # every option's probability

$0.025 per million input tokens. Output is free and there is no minimum per call. Input is everything you send and everything the model reads to answer it, at every step:

Every answer says what it counted (usage.input_tokens_detail); the console's Usage page shows every call.

Errors are {"error": {"message", "type", "code"}}. Refused calls are never charged.

No text, image, recording or answer is stored, logged or used for training. For each call we keep its time, key, kind and structure (which inputs, how many questions of which type, reasons on or off), tokens, cost, status and the country it came from, for billing and your call history (the console shows 30 days).