Troubleshooting

Every error is JSON: {"error": {"type": "...", "message": "..."}}. The message says what to fix.

The answer is not what you expected

• A choice picks the least bad option. Add other or none_of_these. Overlapping descriptions split the probability; make options exclusive.
• A score sits in the middle for everything. Levels written as degree words ("moderate", "severe") do not separate; write each level as a situation ("a workaround exists", "no workaround"). A score over a whole document is blunt; ask per-claim Nouls instead.
• A noul is near 0.5. The engine cannot tell from the state. Add the fact that decides it to the state, or to the subject's memory, and ask a narrower question.
• The answer has an unknown_field warning. The request, its memory or a question has a field Jepela does not read, and the answer was made without it; the message names the closest field it reads. A misspelt subject (subjet) gives a decision without memory, so the answer has no memory block; fix the name and send it again. The memory routes warn the same way: a rule with ttl_secs was stored without an expiry (delete it with POST /v1/memory/rules/delete and add it again with ttl_seconds).
• The memory did not change anything. Check memory.lines_used and memory.lines_seen. Zero used means recall found no line sharing words with the state, the subject's names or the question: put the subject's name and your states' vocabulary in the lines, add the names the subject goes by with POST /v1/memory/aliases, or pin a line that must always count. Fewer seen than used means the lines and the state together are longer than the engine reads; each line in memory.hits and memory.rules_fired says with read whether it was read. In an array Jepela leaves out the oldest elements so that the lines are read (reading.turns_dropped), so there fewer seen means the lines and the newest element alone are longer than the engine reads: send shorter lines. lines_dropped counts lines that did not fit in its 64 context lines. Use compare: true to see the two answers side by side, explain: true to see how far each line moved them, and placebo: true to see whether a change came from what the lines say.
• A rule in memory is ignored or backfires. The engine does not compare numbers or read negation: "never 17" makes 17 more likely, "stop below 30" is answered by its words. Remove options that must never be chosen in code (an exclusion rule, or leave them out of the question), write what to do, and put conditions in a rule with when so they are checked in code (Rules in memory). The warnings say when a line has this problem.
• The answer ignores the end of a long state. The engine read only the start: usage.state_truncated is true and there is a state_cut warning. Send the part that matters first, or "windows": true. An array is the other way round: the engine reads its newest elements and cuts the oldest, so put what matters last; with memory or computed lines, Jepela leaves out the oldest elements so that the lines are read, and reading.turns_dropped and the state_cut warning say how many. With windows, an array longer than 16 windows is read from its end, so it also loses its oldest elements, never its newest.
• A large choice picks strangely. Above 20 options the answer is matched by default ("method": "match" in the answer): give each option a short description of what it means, since an option without one is matched by its name alone. With "method": "engine" it comes from two rounds (rounds: 2) in groups that follow the order you wrote, and "robust": true does not apply; a hierarchy of smaller choices is usually clearer. Up to 20 options, if the answer changes when you reorder them, use "robust": true.
• A request takes longer than usual. The engine loads its three checkpoints when it starts (load_seconds in the engine's /health); only a tenant's fine-tuned checkpoint loads on its first request. After that requests take tens of milliseconds; usage.engine_ms says where the time goes.
• The memory changed an answer it should not have. Lines are recalled by shared words, so a line about another topic can come along. Try "memory": {"focus": true}, which keeps only lines that share a word with the state beyond the subject's name, or raise min_share.