https://www.tarkova.com/blog/Practical reads on semantic caching, agent memory, LLM cost and LLM classification.en https://www.tarkova.com/blog/write-llm-classification-instructions/https://www.tarkova.com/blog/write-llm-classification-instructions/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesWording lessons from Curva's shipped recipes: judge by need, define the no case, carve out false positives, and freeze wording before collecting labels. https://www.tarkova.com/blog/switch-from-jev/https://www.tarkova.com/blog/switch-from-jev/Sat, 03 Oct 2026 00:00:00 GMTcurva vs jevMove a Jev integration to Curva: /v1/systemone, the criteria shape, two environment variables, a shadow test, and the differences to expect. https://www.tarkova.com/blog/support-ticket-triage-ai/https://www.tarkova.com/blog/support-ticket-triage-ai/Sat, 03 Oct 2026 00:00:00 GMTcurva use casesThe support-triage recipe end to end: route confident tickets, queue unsure ones, answer obvious ones with rules, and learn from agents' corrections. https://www.tarkova.com/blog/stop-parsing-llm-text/https://www.tarkova.com/blog/stop-parsing-llm-text/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsDecisions need types and probabilities, not prose. Why parsing a model's sentence fails quietly, and what a typed, calibrated decision gives you instead. https://www.tarkova.com/blog/state-vs-instructions-llm/https://www.tarkova.com/blog/state-vs-instructions-llm/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesWhat belongs in the state and what belongs in the question when you ask an LLM for a decision, and which numbers to compute in code before you ask. https://www.tarkova.com/blog/send-feedback-llm-classifier-python/https://www.tarkova.com/blog/send-feedback-llm-classifier-python/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesSend true labels to an LLM classifier from Python: store the decision id, use the right label per question type, and avoid 404 and 422 errors. https://www.tarkova.com/blog/screenshot-ticket-classification/https://www.tarkova.com/blog/screenshot-ticket-classification/Sat, 03 Oct 2026 00:00:00 GMTcurva use casesSend a ticket's screenshot to a vision model next to its text: is an error visible, which part of the app. Images inline, up to 8 per request. https://www.tarkova.com/blog/reliability-diagram-llm/https://www.tarkova.com/blog/reliability-diagram-llm/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsRead a reliability diagram for an LLM classifier from Curva's reliability bins or dashboard, and spot overconfident, underconfident and leaning models. https://www.tarkova.com/blog/refund-request-detection/https://www.tarkova.com/blog/refund-request-detection/Sat, 03 Oct 2026 00:00:00 GMTcurva use casesA yes/no question that separates "I was charged twice" from "please refund me": the recipe's wording, the probability it returns and how to calibrate it. https://www.tarkova.com/blog/python-plain-vs-full-llm-answers/https://www.tarkova.com/blog/python-plain-vs-full-llm-answers/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesd.refund is True whenever P(yes) is at least 0.5. When plain answers are fine, when to read the full answer, and the id and model name clash. https://www.tarkova.com/blog/python-llm-api-error-handling/https://www.tarkova.com/blog/python-llm-api-error-handling/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesMap each CurvaError subclass to the fix: 422 names the bad question, 429 carries retry_after, 502 splits model_error from model_unavailable. https://www.tarkova.com/blog/product-tagging-ai/https://www.tarkova.com/blog/product-tagging-ai/Sat, 03 Oct 2026 00:00:00 GMTcurva use casesTag products with a Multi question: an independent probability per tag, a threshold you tune, and up to 20 tags per question across a whole catalog. https://www.tarkova.com/blog/pin-llm-prompt-version/https://www.tarkova.com/blog/pin-llm-prompt-version/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesA pinned config freezes the prompt template, mode, debias and model so upgrades don't move your answers. The pins, curva-latest, and when to recalibrate. https://www.tarkova.com/blog/ollama-llm-classification/https://www.tarkova.com/blog/ollama-llm-classification/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesRun typed LLM classification on your own machine with Ollama: no key, no network hop, privacy strict allowed, and a cost of $0 per call. https://www.tarkova.com/blog/none-of-the-above-llm/https://www.tarkova.com/blog/none-of-the-above-llm/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsA classifier forced to pick a label will, confidently. How a none_of_these option works, why it is on by default, and how to route it to a person. https://www.tarkova.com/blog/logprobs-vs-verbal-confidence/https://www.tarkova.com/blog/logprobs-vs-verbal-confidence/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsToken logprobs or a stated probability per label? How each mode works, which one auto picks, and the probe results that changed Curva's default model. https://www.tarkova.com/blog/llm-yes-no-probability/https://www.tarkova.com/blog/llm-yes-no-probability/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesA Noul question returns one number, P(yes), asked as a normalised two-option choice so P(x) and P(not x) agree. How to word it, threshold it and label it. https://www.tarkova.com/blog/llm-image-classification-python/https://www.tarkova.com/blog/llm-image-classification-python/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesClassify images with a vision LLM from Python: paths, bytes or URLs in images=, typed answers with probabilities, and the privacy and size limits. https://www.tarkova.com/blog/llm-decision-tree-one-request/https://www.tarkova.com/blog/llm-decision-tree-one-request/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesBranch on an earlier LLM answer inside one request: @key in when reads an answer, follow-ups run only on the branch taken, and skips cascade to dependents. https://www.tarkova.com/blog/llm-classification-many-classes/https://www.tarkova.com/blog/llm-classification-many-classes/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesHow an LLM Choice question behaves as labels grow: descriptions, the escape option, and the switch to verbal mode with the top 5 labels past 20 options. https://www.tarkova.com/blog/llm-calibration-held-out-gate/https://www.tarkova.com/blog/llm-calibration-held-out-gate/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsCalibration can make LLM probabilities worse. How a five-fold held-out gate on log-loss and Brier decides when to apply it, with a real CommonsenseQA case. https://www.tarkova.com/blog/llm-calibration-explained/https://www.tarkova.com/blog/llm-calibration-explained/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsWhat LLM calibration means, how ECE and Brier measure it, and how Curva fits a calibrator from 30 feedback labels, with held-out before and after numbers. https://www.tarkova.com/blog/llm-audit-log/https://www.tarkova.com/blog/llm-audit-log/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesWhat Curva's audit log records for every LLM decision, what it never keeps (the state, the images, provider errors), and how to page through it. https://www.tarkova.com/blog/install-curva/https://www.tarkova.com/blog/install-curva/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesThree ways to install Curva, what each gives you, which to pick, and three checks that prove it runs: curva --version, GET /health and a first decision. https://www.tarkova.com/blog/how-curva-works/https://www.tarkova.com/blog/how-curva-works/Sat, 03 Oct 2026 00:00:00 GMTcurva engineeringThe path of one Curva request: rules, cache, fenced prompt, two option orders, logprobs or verbal reading, calibration, coverage set, abstain and audit. https://www.tarkova.com/blog/few-shot-examples-llm-classification/https://www.tarkova.com/blog/few-shot-examples-llm-classification/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesAdd up to 10 labeled examples per question, written like feedback labels and remapped when debiasing reverses the options. Which ones to pick. https://www.tarkova.com/blog/fair-coin-llm-test/https://www.tarkova.com/blog/fair-coin-llm-test/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsA coin flip has one honest probability. One published probe got 0.92 from a typed-decision model; Curva gave 0.494. What the test shows and its limits. https://www.tarkova.com/blog/extract-numbers-from-text-llm/https://www.tarkova.com/blog/extract-numbers-from-text-llm/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesNumber and Integer questions return a typed value checked against min and max, with a confidence. What a bad reply becomes, and why sums stay in code. https://www.tarkova.com/blog/expected-calibration-error-llm/https://www.tarkova.com/blog/expected-calibration-error-llm/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsHow expected calibration error is computed, worked by hand on 100 answers, and why an LLM can match on accuracy yet lose on ECE, with measured numbers. https://www.tarkova.com/blog/display-name-spoofing-detection/https://www.tarkova.com/blog/display-name-spoofing-detection/Sat, 03 Oct 2026 00:00:00 GMTcurva use casesThe sender_mismatch check: the display name claims one organisation but the address or reply-to is another domain. Why it is a separate question. https://www.tarkova.com/blog/customer-frustration-score/https://www.tarkova.com/blog/customer-frustration-score/Sat, 03 Oct 2026 00:00:00 GMTcurva use casesUse an ordered Score instead of positive or negative: define the levels, read the expected level (0.65 sits between calm and annoyed), and act on it. https://www.tarkova.com/blog/curva-troubleshooting-guide/https://www.tarkova.com/blog/curva-troubleshooting-guide/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesEvery Curva error in one place: status, type, the SDK exception it raises, the usual causes and the fix, plus surprises that are not errors at all. https://www.tarkova.com/blog/curva-local-python-server/https://www.tarkova.com/blog/curva-local-python-server/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesThree ways to get a Curva client in Python: module-level decide, curva.local() with its own server, or Curva() for a shared one. What each does. https://www.tarkova.com/blog/content-moderation-api/https://www.tarkova.com/blog/content-moderation-api/Sat, 03 Oct 2026 00:00:00 GMTcurva use casesThe content-moderation recipe: a violation category, a severity score, checks for minors and targeted harassment, and a review queue at 0.85. https://www.tarkova.com/blog/conformal-prediction-llm/https://www.tarkova.com/blog/conformal-prediction-llm/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsSet coverage=0.95 and each answer carries a set of labels holding the right one at least 95% of the time. How split conformal works and what it needs. https://www.tarkova.com/blog/conformal-prediction-assumptions/https://www.tarkova.com/blog/conformal-prediction-assumptions/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsConformal prediction sets need labels that look like your traffic. How review-only labels and shifting inputs break coverage, and how to catch it early. https://www.tarkova.com/blog/conditional-llm-questions-when/https://www.tarkova.com/blog/conditional-llm-questions-when/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesAdd when to a question and it is asked only when the state matches. Skipped questions are never sent, stored or paid for, and calibration survives. https://www.tarkova.com/blog/check-llm-logprobs-support/https://www.tarkova.com/blog/check-llm-logprobs-support/Sat, 03 Oct 2026 00:00:00 GMTcurva guidescurva spike tests whether a model returns usable label probabilities from logprobs. Run it on free models, a named provider or your own served model. https://www.tarkova.com/blog/chargeback-reason-classification/https://www.tarkova.com/blog/chargeback-reason-classification/Sat, 03 Oct 2026 00:00:00 GMTcurva use casesA Choice over your chargeback reason codes with descriptions per code, the escape option on, and abstain so disputed cases reach a person. https://www.tarkova.com/blog/bias-scaling-llm-calibration/https://www.tarkova.com/blog/bias-scaling-llm-calibration/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsWhen an LLM favours one class, temperature scaling cannot reorder answers. Bias scaling adds a per-answer offset; measured on AITA and Yelp, held out. https://www.tarkova.com/blog/batch-llm-classification-jsonl/https://www.tarkova.com/blog/batch-llm-classification-jsonl/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesClassify thousands of records from the command line with any LLM: typed answers per line, rate limits respected, and a rerun resumes where it stopped. https://www.tarkova.com/blog/banking-intent-classification/https://www.tarkova.com/blog/banking-intent-classification/Sat, 03 Oct 2026 00:00:00 GMTcurva use casesBuild a 77-intent banking classifier with one Choice: why over 20 options Curva switches to verbal mode, and what Gemini scored on BANKING77 (n = 120). https://www.tarkova.com/blog/abuse-report-triage/https://www.tarkova.com/blog/abuse-report-triage/Sat, 03 Oct 2026 00:00:00 GMTcurva use casesRoute abuse reports by type and severity with min_confidence, so unsure reports go straight to a person and their verdicts calibrate the model. https://www.tarkova.com/blog/abstain-vs-conformal-prediction/https://www.tarkova.com/blog/abstain-vs-conformal-prediction/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsAbstain checks the top answer; a conformal set checks every option. A decision table by question type and the routing rule that combines both. https://www.tarkova.com/blog/reduce-llm-classification-cost/https://www.tarkova.com/blog/reduce-llm-classification-cost/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesEvery way to cut the cost of LLM classification in one place: rules, when, the cache, debias auto, cascades, prompt caching, local models and spend caps. https://www.tarkova.com/blog/python-llm-classification/https://www.tarkova.com/blog/python-llm-classification/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesA start-to-finish Python tutorial: typed LLM labels with a probability each, abstain, feedback, calibration reports, async batches and errors. https://www.tarkova.com/blog/phishing-detection-llm/https://www.tarkova.com/blog/phishing-detection-llm/Sat, 03 Oct 2026 00:00:00 GMTcurva use casesPhishing detection with an LLM that returns P(phishing), tactics, a risk level and a sender check, measured on PhishNChips with its limits. https://www.tarkova.com/blog/n8n-ai-routing/https://www.tarkova.com/blog/n8n-ai-routing/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesRoute n8n items with an LLM: one branch per answer, a Needs review branch for unsure ones, and a Feedback node that calibrates on your data. https://www.tarkova.com/blog/llm-position-bias/https://www.tarkova.com/blog/llm-position-bias/Sat, 03 Oct 2026 00:00:00 GMTcurva conceptsLLMs favour options by their position in a list. How asking in two orders and averaging cancels it, what it costs, and when debias auto skips a call. https://www.tarkova.com/blog/llm-decisions-over-http/https://www.tarkova.com/blog/llm-decisions-over-http/Sat, 03 Oct 2026 00:00:00 GMTcurva guidesGet typed LLM decisions from any language or no-code tool with two HTTP calls: decide and feedback. Request, response, routing order, errors.