# FAQ

**Is Jepela a chat model?** No. It never generates text. You send a state and typed questions; you get typed answers with probabilities. Use it inside software, for judgments the code then acts on.

**What is a "state"?** The thing to judge: a message, a record, a document, a scene. A string, an object with named fields, or an array of strings. Text only.

**What is a "subject"?** Whatever the decision is about and worth remembering things for: a customer, a user, a player, a device, a case, a project. Name it on a request and Jepela recalls that subject's memory lines and shows which ones it used.

**How many questions can I ask in one request?** Up to 400 engine questions: one per question, one per option order of a robust choice, and for a choice over 20 options read by the engine one per group of 20 plus one final round (a matched choice is not an engine question). Each is answered on its own against the same state, and adding questions does not change the other answers. The body is capped at 1 MB. Fifty questions in one request is tested.

**How many options can a choice have?** 1 to 255. Up to 20 are read by the engine in one pass. Above that Jepela matches by default: the state and each option's description as vectors, the closest wins (65.2% right against the engine's 62.2% on 77 bank intents, in 2 ms instead of 209 ms). With `"method": "engine"` the engine reads them in two rounds instead (groups, then the group winners), measured right on 8 of 8 at 40, 80 and 255 options. A score has 2 to 10 levels.

**What do I pay for?** Input tokens, $0.042 per million, the same way and price as Jev; output is free. `usage.input_tokens` is the engine's own count. Memory writes are free; recalled lines count as input. A matched choice pays only for the state and memory it read, not its options. Prices are in `GET /v1/usage`.

**What does confidence mean?** How concentrated the probabilities are: near 1 when one option dominates, low when they are spread. It is the engine's estimate; whether 0.9 means "right nine times in ten" for your data is something to measure on your own labelled examples, and Jepela can fit it to them (Quality). Use it as a gate: act above a threshold, confirm in the middle, hand over below.

**Is the engine deterministic?** The same request to the same checkpoint gives the same answer on the same machine. Different hardware or checkpoint versions can round differently.

**Why did a rule in memory not work?** Three known reasons, all measured: the engine does not compare numbers, it does not read negation, and lines move answers by being there, in either direction. Put conditions in code (a memory rule with `when`, or `derive`), remove options that must never be chosen with an exclusion rule, pin what must always count, keep memories short, and check with `explain` and `placebo`. See Rules in memory.

**Why did a fact not get recalled?** Recall matches words and short phrases between the lines and the state, the subject's names and aliases, and the question's own words. Put the vocabulary of your states into the lines, add other names the subject goes by (`POST /v1/memory/aliases`), or pin a line that must count at every decision.

**Can I see what the memory changed?** Yes: `"memory": {"compare": true}` returns the answer without memory and `changes` per question. It doubles the engine work for that request and is not billed.

**Can I delete everything about a subject?** `POST /v1/memory/delete` removes the subject's memory and rules from RAM and disk at once. The tenant's ledger keeps the subject id on each priced request, and batches and golden cases keep the requests that named it until you delete them. `forget` removes one concept and verifies the removal against recall.

**Which languages?** English first. `model: jepela` routes by language and picks the multilingual checkpoint when the state is not English; it is less measured. Do not name `jepela-english` for other languages: its authors report that it collapses outside English while staying confident.

**Can I send images?** No. Describe the image in the state.

**What happens when the engine is down?** 502 (`the engine is off` when the GPU is switched off) and nothing is charged. Jepela never invents an answer. The SDKs retry with backoff.

**Can I train a model on my own cases?** Yes: fine-tuning trains a model of your own on your golden cases (at least 20). It is used only if it beats the standard engine on cases it never saw and still listens to memory, and only your account can use it. Free during the preview; see Fine-tuning.

**Memory or fine-tuning?** Facts about a person or an account go in memory (they can be forgotten at once); how you decide across all cases is what training teaches. The two work together.

**Is there a JavaScript SDK?** Yes: `sdk/typescript`, a TypeScript client with no dependencies that Node 22.18 or newer runs directly; see JavaScript and TypeScript. The same page shows the few lines of `fetch` you need without it.

**Can Claude use Jepela?** Yes: an MCP server and a Claude Code plugin; see Claude and MCP.