Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
docs(integrations): the lang_guess table must report what current main routes
Aashishb59c32093ae1
diff --git a/docs/langchain.md b/docs/langchain.mdindex 9c6f6a34..7ae5d0c4 100644--- a/docs/langchain.md+++ b/docs/langchain.md@@ -501,22 +501,29 @@ Both are per node, like the budgets, and unset means unsent: leave one out and t`Router(task=...)` / `Router(lang_guess=...)` stays in charge rather than being overwritten with`None`. `batch()` sends them with every request in the batch, `abatch()` on the same path.-**What the hint moves, measured.** Sixteen German support queries through `Router._route` on-`6d942c9` -- routing only, so no checkpoint is loaded and no answer is taken -- counted by which-checkpoint the decision lands on:+**What the hint moves, measured.** Sixteen German support queries through `Router.route` on+`6880462` (0.4.2) -- routing only, so no checkpoint is loaded and no answer is taken -- counted by+which checkpoint the decision lands on:| German queries | no hint | `lang_guess="de"` ||---|---|---|-| 8 written without umlauts or `ß` | 6 answered on `english` | 8 on `multilingual` |-| 8 carrying umlauts or `ß` | 2 answered on `english` | 8 on `multilingual` |--Detection is a good guesser when the text gives it letters to work with; `Zahlung fehlgeschlagen`-brings it to its stated abstention (`"Latin script, language not identified and no non-English-letters; using default (english)"`), and `Was kostet die MwSt?` to a wrong read (`"English Latin-text"`). A caller that knows the language from a user id or an `Accept-Language` header knows better-than that, which is the whole argument for the hint. What is *not* claimed here is what the two-checkpoints then score on those queries -- that needs weights and a labelled set, and this table is-about which one answers, not how well.+| 8 written without umlauts or `ß` | 1 read as English | 0 read as English |+| 8 carrying umlauts or `ß` | 0 read as English | 0 read as English |++The one miss is `Was kostet die MwSt?`, which detection reads as `"English Latin text"`: an+abbreviation-heavy short query with no diacritics and no stopword signal. Diacritics settle it on+their own, so the eight `umlauts or ß` queries all leave `english` without any help from the hint.+Plain German prose that detection cannot place does not reach the English checkpoint either any+more -- an undecided Latin script now abstains to the multilingual one (`"Latin script, language+not identified and no non-English letters; using default (multilingual)"`, asserted in+`tests/test_router.py`), so the residual gap the hint closes is the false-English read, not the+undecided one. A caller that knows the language from a user id or an `Accept-Language` header still+knows better than any guess, which is the argument for the hint. What is *not* claimed here is what+the two checkpoints then score on those queries -- that needs weights and a labelled set, and this+table is about which one answers, not how well.++`tests/test_langchain.py` re-runs these sixteen queries against the live router and asserts the+hint half of the table, so the numbers above cannot quietly go stale behind this paragraph.**A recorded `before`.** On `6d942c9`, in real (installed `langchain-core`) mode, a node built withboth hints and a budget, answering through a runner that prints what `predict` received:diff --git a/docs/llamaindex.md b/docs/llamaindex.mdindex 1c5e2a24..b1aaf8e6 100644--- a/docs/llamaindex.md+++ b/docs/llamaindex.md@@ -217,11 +217,13 @@ workflow rather than waiting for auto task detection to recognise the question itakes a language code -- or a callable of the query being selected against -- from something thatalready knows it, consulted after an explicit `lang` and before core's own script detection; anynon-English code routes to the multilingual checkpoint. A German-documented index is the case the-hint is for: measured on sixteen German support queries, detection leaves 6 of the 8 written without-umlauts and 2 of the 8 carrying them on the `english` checkpoint, while `lang_guess="de"` moves all-16 to `multilingual`. See the [LangChain-integration](langchain.md#9-which-checkpoint-answers) for that table and its-reason strings.+hint is for, but the gap is narrower than it looks: measured on sixteen German support queries+through `Router.route` on `6880462`, one of the eight written without umlauts reaches `english`+(`Was kostet die MwSt?`, read as "English Latin text"), and none of the eight carrying umlauts or+`ß` does -- diacritics settle it, and an undecided Latin script now abstains to `multilingual`+rather than to `english`. `lang_guess="de"` takes that last one off `english` too. See the [LangChain+integration](langchain.md#9-which-checkpoint-answers) for that table, its reason strings, and the+check that re-runs the sixteen queries.```pythonselector = LayaSingleSelector(diff --git a/tests/test_langchain.py b/tests/test_langchain.pyindex 4d60d800..5b1d146f 100644--- a/tests/test_langchain.py+++ b/tests/test_langchain.py@@ -1627,6 +1627,42 @@ except ValueError as exc:"RouteIgnoringHintsRouter.route" in str(exc), str(exc))+# The `docs/langchain.md` "What the hint moves, measured" table, re-run against the live router.+# Routing only: `route` decides without loading a checkpoint, so this costs no weights and no+# forward pass. The assertion is the guarantee the hint gives -- a caller that says "de" gets the+# multilingual checkpoint for every one of these. The no-hint column is printed, not pinned: it is+# the current detection heuristic's behaviour, and a run that fixes #54 should move the print, not+# turn the suite red.+from laya.router import Router as _RoutingRouter # noqa: E402++_DE_NO_DIACRITICS = [+ "Zahlung fehlgeschlagen", "Wo finde ich meine Rechnung", "Ich kann mich nicht anmelden",+ "Die App startet nicht", "Wie aendere ich meine Adresse", "Bestellung noch nicht angekommen",+ "Ich brauche eine Kuendigung", "Was kostet die MwSt?",+]+_DE_DIACRITIC = [+ "Rechnung für die MwSt stimmt nicht", "Paßwort zurücksetzen geht nicht",+ "Größe des Pakets passt nicht", "Kündigung per E-Mail senden",+ "Gehälter ohne Umlaute im Profil", "Wie viel kostet der Versand heute",+ "Ich möchte meine Adresse ändern", "Wo muß ich das Formular einreichen",+]+_DE_ALL = _DE_NO_DIACRITICS + _DE_DIACRITIC+_Q_GUESS = {"a": {"type": "noul", "instructions": "Is `text` a support question?"}}+_rt = _RoutingRouter()+++def _on_english(texts, **kw):+ return [t for t in texts if _rt.route({"text": t}, _Q_GUESS, **kw)["model"] == "english"]+++print("\n lang_guess table, this run (queries read as English):")+for _label, _qs in (("without umlauts or ß", _DE_NO_DIACRITICS), ("with umlauts or ß", _DE_DIACRITIC)):+ print(" %-24s no hint %d, lang_guess=\"de\" %d %s"+ % (_label, len(_on_english(_qs)), len(_on_english(_qs, lang_guess="de")), _on_english(_qs)))+check("routing/with the de hint no German query is read as English",+ _on_english(_DE_ALL, lang_guess="de"), [])++# --------------------------------------------------------------- Summaryprint(f"PASS: {len(PASS)}")print(f"FAIL: {len(FAIL)}")