State

What Jev reads. Strings, objects and arrays, paths into them, and structured questions.

Follows docs.typesafe.ai/concepts/state and the advanced-structure section of primitives.

State is the material you put in front of the model before asking. The docs' image: what you would hand a panel of experts before asking for their judgement. One request evaluates one state against every question in it.

Three shapes

shape	use it for	example
string	one message, passage or article	"My card was charged twice."
object	named fields, related records, application state	{"message": "...", "order_id": "A-104"}
array	a sequence of messages or records	["Hi", "My customer number is TS1337.", "My card was charged twice."]

Use an object for most requests. Field names carry meaning to the model, and a question can point into the object with a backticked path:

state = {
    "ticket": {"subject": "Duplicate charge",
               "messages": [{"from": "customer", "text": "I was charged twice for A-104."},
                            {"from": "support", "text": "We are checking the charges."}]},
    "order": {"id": "A-104", "charges": [{"amount_usd": 49}, {"amount_usd": 49}]},
    "refund_policy": "Duplicate charges are eligible for a refund.",
}
questions = {
    "asks_refund": "Does `ticket.messages[0].text` ask for a refund?",
    "policy_covers": "Does `refund_policy` cover what `ticket.messages[0].text` describes?",
}

Text only at jev-1.13: strings, objects and arrays of text. No images, audio or video.

Keep content and questions apart

The state holds the facts: the message, the order, the policy. The questions hold the judgements: did they ask, does the policy allow. Do not fold the policy into the instructions of a question; put it in the state where every question can read it, and ask about it by path.

Structured instructions and criteria

Instructions and criteria may be JSON values rather than strings when the question has structure of its own: a rubric with named parts, a set of examples, a list of fields to compare. The SDK serialises them; the model reads them as part of the question. jev_ask passes them through unchanged.

How much state

Two limits from the API: 32,000 tokens for the state plus the longest question, 64,000 per request. The client estimates and refuses before sending (What it costs). Well under those limits, less is still better: the jaggedness page reports that large irrelevant state moves answers, so the many-item tools in this package send one item per request rather than a list, and render_state truncates a single item at 12,000 characters with a marker. Filter first, then ask.

Edit page