smallm 0.5.0

package/dist/cli.d.ts

@@ -0,0 +1,32 @@
+#!/usr/bin/env node
+import { ModelQuery, ModelMatch } from "./types";
+/**
+ * Thin CLI wrapper around findModels(). Per the v0.5 guide's Do's/Don'ts:
+ * this file must contain NO scoring, filtering, or provider logic — only
+ * flag parsing (turning argv into a ModelQuery) and output formatting
+ * (turning a ModelMatch[] into text). All actual decisions — including
+ * input validation — are made by the library itself; this file just calls
+ * findModels() and displays whatever it returns or throws.
+ */
+/** Thrown for usage problems (unknown flag, unknown command, --help). Caught in main(), never process.exit()'d from inside parseArgs itself — keeps parseArgs testable as a pure function. */
+export declare class CliUsageError extends Error {
+    readonly exitCode: number;
+    constructor(message: string, exitCode?: number);
+}
+export interface ParsedCli {
+    query: ModelQuery;
+    json: boolean;
+    showHelp: boolean;
+}
+export declare const USAGE = "Usage: smallm find [options]\n\nRequired:\n  --task <task>                Task type (chat, summarize, classify, extract, code, translate, or any custom string)\n  --context <tokens>           Required context window, in tokens\n\nOptional (map 1:1 to ModelQuery fields):\n  --max-params <billions>      Hard filter, e.g. 7 = \"under 7B\"\n  --domain <domain>            Preferred domain tag, used in scoring only\n  --hardware <cpu|gpu-low|gpu-high>   Hardware hint (original string form)\n  --hardware-type <cpu|gpu>    Hardware type (structured form \u2014 use with --hardware-vram)\n  --hardware-vram <gb>         VRAM budget in GB (structured form only)\n  --max-latency <ms>           Enforced only when benchmark data exists for the (model, hardware) pair\n  --scoring-mode <rule|embedding|hybrid>   Scoring strategy (default: rule)\n  --providers <list>           Comma-separated, e.g. huggingface,ollama (default: huggingface)\n  --limit <n>                  Max results (default: 5)\n  --cache-dir <path>           File cache directory\n  --cache-ttl <ms>             File cache TTL in ms\n\nOutput:\n  --json                       Print raw ModelMatch[] JSON instead of a table\n\nExamples:\n  npx smallm find --task summarize --context 4096\n  npx smallm find --task chat --context 8192 --max-params 7 --json\n  npx smallm find --task chat --context 4096 --providers huggingface,ollama --hardware-type gpu --hardware-vram 8\n";
+/** Pure parser: argv (already sliced past "find") -> a ModelQuery + output mode. Never touches process.exit or the network. */
+export declare function parseArgs(argv: string[]): ParsedCli;
+/** Pure formatter: ModelMatch[] -> a human-readable table string. */
+export declare function formatTable(results: ModelMatch[]): string;
+/**
+ * Entry point. Only place in this file that touches process.exit/console —
+ * everything else (parseArgs, formatTable) is a pure, independently
+ * testable function.
+ */
+export declare function main(argv: string[]): Promise<number>;
+//# sourceMappingURL=cli.d.ts.map

package/dist/cli.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"cli.d.ts","sourceRoot":"","sources":["../src/cli.ts"],"names":[],"mappings":";AAEA,OAAO,EAAE,UAAU,EAAE,UAAU,EAAE,MAAM,SAAS,CAAC;AAEjD;;;;;;;GAOG;AAEH,8LAA8L;AAC9L,qBAAa,aAAc,SAAQ,KAAK;aACO,QAAQ,EAAE,MAAM;IAA7D,YAAY,OAAO,EAAE,MAAM,EAAkB,QAAQ,GAAE,MAAU,EAGhE;CACF;AAED,MAAM,WAAW,SAAS;IACxB,KAAK,EAAE,UAAU,CAAC;IAClB,IAAI,EAAE,OAAO,CAAC;IACd,QAAQ,EAAE,OAAO,CAAC;CACnB;AAED,eAAO,MAAM,KAAK,i6CA0BjB,CAAC;AAEF,+HAA+H;AAC/H,wBAAgB,SAAS,CAAC,IAAI,EAAE,MAAM,EAAE,GAAG,SAAS,CAoInD;AAMD,qEAAqE;AACrE,wBAAgB,WAAW,CAAC,OAAO,EAAE,UAAU,EAAE,GAAG,MAAM,CAmBzD;AAED;;;;GAIG;AACH,wBAAsB,IAAI,CAAC,IAAI,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,MAAM,CAAC,CA0C1D"}

package/dist/cli.js

@@ -0,0 +1,256 @@
+#!/usr/bin/env node
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.USAGE = exports.CliUsageError = void 0;
+exports.parseArgs = parseArgs;
+exports.formatTable = formatTable;
+exports.main = main;
+const index_1 = require("./index");
+/**
+ * Thin CLI wrapper around findModels(). Per the v0.5 guide's Do's/Don'ts:
+ * this file must contain NO scoring, filtering, or provider logic — only
+ * flag parsing (turning argv into a ModelQuery) and output formatting
+ * (turning a ModelMatch[] into text). All actual decisions — including
+ * input validation — are made by the library itself; this file just calls
+ * findModels() and displays whatever it returns or throws.
+ */
+/** Thrown for usage problems (unknown flag, unknown command, --help). Caught in main(), never process.exit()'d from inside parseArgs itself — keeps parseArgs testable as a pure function. */
+class CliUsageError extends Error {
+    constructor(message, exitCode = 1) {
+        super(message);
+        this.exitCode = exitCode;
+        this.name = "CliUsageError";
+    }
+}
+exports.CliUsageError = CliUsageError;
+exports.USAGE = `Usage: smallm find [options]
+
+Required:
+  --task <task>                Task type (chat, summarize, classify, extract, code, translate, or any custom string)
+  --context <tokens>           Required context window, in tokens
+
+Optional (map 1:1 to ModelQuery fields):
+  --max-params <billions>      Hard filter, e.g. 7 = "under 7B"
+  --domain <domain>            Preferred domain tag, used in scoring only
+  --hardware <cpu|gpu-low|gpu-high>   Hardware hint (original string form)
+  --hardware-type <cpu|gpu>    Hardware type (structured form — use with --hardware-vram)
+  --hardware-vram <gb>         VRAM budget in GB (structured form only)
+  --max-latency <ms>           Enforced only when benchmark data exists for the (model, hardware) pair
+  --scoring-mode <rule|embedding|hybrid>   Scoring strategy (default: rule)
+  --providers <list>           Comma-separated, e.g. huggingface,ollama (default: huggingface)
+  --limit <n>                  Max results (default: 5)
+  --cache-dir <path>           File cache directory
+  --cache-ttl <ms>             File cache TTL in ms
+
+Output:
+  --json                       Print raw ModelMatch[] JSON instead of a table
+
+Examples:
+  npx smallm find --task summarize --context 4096
+  npx smallm find --task chat --context 8192 --max-params 7 --json
+  npx smallm find --task chat --context 4096 --providers huggingface,ollama --hardware-type gpu --hardware-vram 8
+`;
+/** Pure parser: argv (already sliced past "find") -> a ModelQuery + output mode. Never touches process.exit or the network. */
+function parseArgs(argv) {
+    const query = {};
+    let json = false;
+    let showHelp = false;
+    let hardwareType;
+    let hardwareVram;
+    let hardwareString;
+    let cacheDir;
+    let cacheTtl;
+    const takeValue = (flag, i) => {
+        const value = argv[i + 1];
+        if (value === undefined) {
+            throw new CliUsageError(`Flag ${flag} requires a value`);
+        }
+        return [value, i + 1];
+    };
+    for (let i = 0; i < argv.length; i++) {
+        const arg = argv[i];
+        switch (arg) {
+            case "--task": {
+                const [v, next] = takeValue(arg, i);
+                query.task = v;
+                i = next;
+                break;
+            }
+            case "--context": {
+                const [v, next] = takeValue(arg, i);
+                query.contextLength = Number(v);
+                i = next;
+                break;
+            }
+            case "--max-params": {
+                const [v, next] = takeValue(arg, i);
+                query.maxParamsB = Number(v);
+                i = next;
+                break;
+            }
+            case "--domain": {
+                const [v, next] = takeValue(arg, i);
+                query.domain = v;
+                i = next;
+                break;
+            }
+            case "--hardware": {
+                const [v, next] = takeValue(arg, i);
+                hardwareString = v;
+                i = next;
+                break;
+            }
+            case "--hardware-type": {
+                const [v, next] = takeValue(arg, i);
+                hardwareType = v;
+                i = next;
+                break;
+            }
+            case "--hardware-vram": {
+                const [v, next] = takeValue(arg, i);
+                hardwareVram = Number(v);
+                i = next;
+                break;
+            }
+            case "--max-latency": {
+                const [v, next] = takeValue(arg, i);
+                query.maxLatencyMs = Number(v);
+                i = next;
+                break;
+            }
+            case "--scoring-mode": {
+                const [v, next] = takeValue(arg, i);
+                query.scoringMode = v;
+                i = next;
+                break;
+            }
+            case "--providers": {
+                const [v, next] = takeValue(arg, i);
+                query.providers = v.split(",").map((s) => s.trim());
+                i = next;
+                break;
+            }
+            case "--limit": {
+                const [v, next] = takeValue(arg, i);
+                query.limit = Number(v);
+                i = next;
+                break;
+            }
+            case "--cache-dir": {
+                const [v, next] = takeValue(arg, i);
+                cacheDir = v;
+                i = next;
+                break;
+            }
+            case "--cache-ttl": {
+                const [v, next] = takeValue(arg, i);
+                cacheTtl = Number(v);
+                i = next;
+                break;
+            }
+            case "--json":
+                json = true;
+                break;
+            case "--help":
+            case "-h":
+                showHelp = true;
+                break;
+            default:
+                throw new CliUsageError(`Unknown flag: ${arg}`);
+        }
+    }
+    // Structured hardware form takes precedence if given; otherwise fall back
+    // to the string-enum form. Both map onto the exact same ModelQuery.hardware
+    // field the library already accepts — no new capability invented here.
+    if (hardwareType !== undefined) {
+        query.hardware = {
+            type: hardwareType,
+            ...(hardwareVram !== undefined ? { vramGB: hardwareVram } : {}),
+        };
+    }
+    else if (hardwareString !== undefined) {
+        query.hardware = hardwareString;
+    }
+    if (cacheDir !== undefined || cacheTtl !== undefined) {
+        query.cacheOptions = {
+            ...(cacheDir !== undefined ? { dir: cacheDir } : {}),
+            ...(cacheTtl !== undefined ? { ttlMs: cacheTtl } : {}),
+        };
+    }
+    return { query: query, json, showHelp };
+}
+function truncate(str, max) {
+    return str.length > max ? str.slice(0, max - 1) + "…" : str;
+}
+/** Pure formatter: ModelMatch[] -> a human-readable table string. */
+function formatTable(results) {
+    if (results.length === 0) {
+        return "No matching models found.";
+    }
+    const columns = [
+        { header: "NAME", width: 42, get: (m) => truncate(m.name, 42) },
+        { header: "PROVIDER", width: 12, get: (m) => m.provider },
+        { header: "PARAMS(B)", width: 10, get: (m) => (m.paramsB == null ? "?" : String(m.paramsB)) },
+        { header: "SCORE", width: 6, get: (m) => String(m.score) },
+        { header: "MODE", width: 10, get: (m) => m.scoringMode ?? "rule" },
+        { header: "REASON", width: 50, get: (m) => truncate(m.reasonWhy, 50) },
+    ];
+    const headerRow = columns.map((c) => c.header.padEnd(c.width)).join(" ");
+    const separator = columns.map((c) => "-".repeat(c.width)).join(" ");
+    const rows = results.map((m) => columns.map((c) => c.get(m).padEnd(c.width)).join(" "));
+    return [headerRow, separator, ...rows].join("\n");
+}
+/**
+ * Entry point. Only place in this file that touches process.exit/console —
+ * everything else (parseArgs, formatTable) is a pure, independently
+ * testable function.
+ */
+async function main(argv) {
+    if (argv.length === 0 || argv.includes("--help") || argv.includes("-h")) {
+        console.log(exports.USAGE);
+        return argv.length === 0 ? 1 : 0;
+    }
+    if (argv[0] !== "find") {
+        console.error(`Unknown command: "${argv[0]}". Only "find" is supported.\n`);
+        console.log(exports.USAGE);
+        return 1;
+    }
+    let parsed;
+    try {
+        parsed = parseArgs(argv.slice(1));
+    }
+    catch (err) {
+        if (err instanceof CliUsageError) {
+            console.error(`Error: ${err.message}\n`);
+            console.log(exports.USAGE);
+            return err.exitCode;
+        }
+        throw err;
+    }
+    if (parsed.showHelp) {
+        console.log(exports.USAGE);
+        return 0;
+    }
+    try {
+        // The one line that matters: flags in, library call, nothing invented.
+        const results = await (0, index_1.findModels)(parsed.query);
+        console.log(parsed.json ? JSON.stringify(results, null, 2) : formatTable(results));
+        return 0;
+    }
+    catch (err) {
+        if (err instanceof index_1.SmallmError) {
+            console.error(`Error: ${err.message}`);
+        }
+        else {
+            console.error("Unexpected error:", err);
+        }
+        return 1;
+    }
+}
+/* istanbul ignore next -- exercised via the compiled bin, not unit tests */
+if (require.main === module) {
+    main(process.argv.slice(2)).then((code) => {
+        process.exitCode = code;
+    });
+}
+//# sourceMappingURL=cli.js.map

package/dist/cli.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cli.js","sourceRoot":"","sources":["../src/cli.ts"],"names":[],"mappings":";;;;;;;AACA,mCAAkD;AAGlD;;;;;;;GAOG;AAEH,8LAA8L;AAC9L,mBAA2B,SAAQ,KAAK;IACtC,YAAY,OAAe,EAAkB,QAAQ,GAAW,CAAC;QAC/D,KAAK,CAAC,OAAO,CAAC,CAAC;wBAD4B,QAAQ;QAEnD,IAAI,CAAC,IAAI,GAAG,eAAe,CAAC;IAC9B,CAAC;CACF;;AAQY,QAAA,KAAK,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;CA0BpB,CAAC;AAEF,+HAA+H;AAC/H,mBAA0B,IAAc;IACtC,MAAM,KAAK,GAA4B,EAAE,CAAC;IAC1C,IAAI,IAAI,GAAG,KAAK,CAAC;IACjB,IAAI,QAAQ,GAAG,KAAK,CAAC;IACrB,IAAI,YAAgC,CAAC;IACrC,IAAI,YAAgC,CAAC;IACrC,IAAI,cAAkC,CAAC;IACvC,IAAI,QAA4B,CAAC;IACjC,IAAI,QAA4B,CAAC;IAEjC,MAAM,SAAS,GAAG,CAAC,IAAY,EAAE,CAAS,EAAoB,EAAE;QAC9D,MAAM,KAAK,GAAG,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QAC1B,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;YACxB,MAAM,IAAI,aAAa,CAAC,QAAQ,IAAI,mBAAmB,CAAC,CAAC;QAC3D,CAAC;QACD,OAAO,CAAC,KAAK,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC;IACxB,CAAC,CAAC;IAEF,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACrC,MAAM,GAAG,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC;QAEpB,QAAQ,GAAG,EAAE,CAAC;YACZ,KAAK,QAAQ,EAAE,CAAC;gBACd,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,KAAK,CAAC,IAAI,GAAG,CAAC,CAAC;gBACf,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,WAAW,EAAE,CAAC;gBACjB,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,KAAK,CAAC,aAAa,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;gBAChC,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,cAAc,EAAE,CAAC;gBACpB,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,KAAK,CAAC,UAAU,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;gBAC7B,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,UAAU,EAAE,CAAC;gBAChB,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC;gBACjB,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,YAAY,EAAE,CAAC;gBAClB,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,cAAc,GAAG,CAAC,CAAC;gBACnB,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,iBAAiB,EAAE,CAAC;gBACvB,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,YAAY,GAAG,CAAC,CAAC;gBACjB,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,iBAAiB,EAAE,CAAC;gBACvB,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,YAAY,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;gBACzB,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,eAAe,EAAE,CAAC;gBACrB,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,KAAK,CAAC,YAAY,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;gBAC/B,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,gBAAgB,EAAE,CAAC;gBACtB,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,KAAK,CAAC,WAAW,GAAG,CAAC,CAAC;gBACtB,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,aAAa,EAAE,CAAC;gBACnB,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC;gBACpD,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,SAAS,EAAE,CAAC;gBACf,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,KAAK,CAAC,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;gBACxB,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,aAAa,EAAE,CAAC;gBACnB,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,QAAQ,GAAG,CAAC,CAAC;gBACb,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,aAAa,EAAE,CAAC;gBACnB,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;gBACpC,QAAQ,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;gBACrB,CAAC,GAAG,IAAI,CAAC;gBACT,MAAM;YACR,CAAC;YACD,KAAK,QAAQ;gBACX,IAAI,GAAG,IAAI,CAAC;gBACZ,MAAM;YACR,KAAK,QAAQ,CAAC;YACd,KAAK,IAAI;gBACP,QAAQ,GAAG,IAAI,CAAC;gBAChB,MAAM;YACR;gBACE,MAAM,IAAI,aAAa,CAAC,iBAAiB,GAAG,EAAE,CAAC,CAAC;QACpD,CAAC;IACH,CAAC;IAED,0EAA0E;IAC1E,4EAA4E;IAC5E,uEAAuE;IACvE,IAAI,YAAY,KAAK,SAAS,EAAE,CAAC;QAC/B,KAAK,CAAC,QAAQ,GAAG;YACf,IAAI,EAAE,YAAY;YAClB,GAAG,CAAC,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,YAAY,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;SAChE,CAAC;IACJ,CAAC;SAAM,IAAI,cAAc,KAAK,SAAS,EAAE,CAAC;QACxC,KAAK,CAAC,QAAQ,GAAG,cAAc,CAAC;IAClC,CAAC;IAED,IAAI,QAAQ,KAAK,SAAS,IAAI,QAAQ,KAAK,SAAS,EAAE,CAAC;QACrD,KAAK,CAAC,YAAY,GAAG;YACnB,GAAG,CAAC,QAAQ,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,QAAQ,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACpD,GAAG,CAAC,QAAQ,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;SACvD,CAAC;IACJ,CAAC;IAED,OAAO,EAAE,KAAK,EAAE,KAA8B,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAC;AACnE,CAAC;AAED,SAAS,QAAQ,CAAC,GAAW,EAAE,GAAW;IACxC,OAAO,GAAG,CAAC,MAAM,GAAG,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC;AAC9D,CAAC;AAED,qEAAqE;AACrE,qBAA4B,OAAqB;IAC/C,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACzB,OAAO,2BAA2B,CAAC;IACrC,CAAC;IAED,MAAM,OAAO,GAA6E;QACxF,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,EAAE,EAAE,GAAG,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,EAAE;QAC/D,EAAE,MAAM,EAAE,UAAU,EAAE,KAAK,EAAE,EAAE,EAAE,GAAG,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,EAAE;QACzD,EAAE,MAAM,EAAE,WAAW,EAAE,KAAK,EAAE,EAAE,EAAE,GAAG,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,OAAO,IAAI,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,EAAE;QAC7F,EAAE,MAAM,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,EAAE;QAC1D,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,EAAE,EAAE,GAAG,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,WAAW,IAAI,MAAM,EAAE;QAClE,EAAE,MAAM,EAAE,QAAQ,EAAE,KAAK,EAAE,EAAE,EAAE,GAAG,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,QAAQ,CAAC,CAAC,CAAC,SAAS,EAAE,EAAE,CAAC,EAAE;KACvE,CAAC;IAEF,MAAM,SAAS,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IACzE,MAAM,SAAS,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IACpE,MAAM,IAAI,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC;IAExF,OAAO,CAAC,SAAS,EAAE,SAAS,EAAE,GAAG,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AACpD,CAAC;AAED;;;;GAIG;AACI,KAAK,eAAe,IAAc;IACvC,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC,IAAI,IAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,IAAI,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,EAAE,CAAC;QACxE,OAAO,CAAC,GAAG,CAAC,QAAA,KAAK,CAAC,CAAC;QACnB,OAAO,IAAI,CAAC,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACnC,CAAC;IAED,IAAI,IAAI,CAAC,CAAC,CAAC,KAAK,MAAM,EAAE,CAAC;QACvB,OAAO,CAAC,KAAK,CAAC,qBAAqB,IAAI,CAAC,CAAC,CAAC,gCAAgC,CAAC,CAAC;QAC5E,OAAO,CAAC,GAAG,CAAC,QAAA,KAAK,CAAC,CAAC;QACnB,OAAO,CAAC,CAAC;IACX,CAAC;IAED,IAAI,MAAiB,CAAC;IACtB,IAAI,CAAC;QACH,MAAM,GAAG,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;IACpC,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACb,IAAI,GAAG,YAAY,aAAa,EAAE,CAAC;YACjC,OAAO,CAAC,KAAK,CAAC,UAAU,GAAG,CAAC,OAAO,IAAI,CAAC,CAAC;YACzC,OAAO,CAAC,GAAG,CAAC,QAAA,KAAK,CAAC,CAAC;YACnB,OAAO,GAAG,CAAC,QAAQ,CAAC;QACtB,CAAC;QACD,MAAM,GAAG,CAAC;IACZ,CAAC;IAED,IAAI,MAAM,CAAC,QAAQ,EAAE,CAAC;QACpB,OAAO,CAAC,GAAG,CAAC,QAAA,KAAK,CAAC,CAAC;QACnB,OAAO,CAAC,CAAC;IACX,CAAC;IAED,IAAI,CAAC;QACH,uEAAuE;QACvE,MAAM,OAAO,GAAG,MAAM,IAAA,kBAAU,EAAC,MAAM,CAAC,KAAK,CAAC,CAAC;QAC/C,OAAO,CAAC,GAAG,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC,CAAC;QACnF,OAAO,CAAC,CAAC;IACX,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACb,IAAI,GAAG,YAAY,mBAAW,EAAE,CAAC;YAC/B,OAAO,CAAC,KAAK,CAAC,UAAU,GAAG,CAAC,OAAO,EAAE,CAAC,CAAC;QACzC,CAAC;aAAM,CAAC;YACN,OAAO,CAAC,KAAK,CAAC,mBAAmB,EAAE,GAAG,CAAC,CAAC;QAC1C,CAAC;QACD,OAAO,CAAC,CAAC;IACX,CAAC;AACH,CAAC;AAED,4EAA4E;AAC5E,IAAI,OAAO,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;IAC5B,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE;QACxC,OAAO,CAAC,QAAQ,GAAG,IAAI,CAAC;IAC1B,CAAC,CAAC,CAAC;AACL,CAAC"}

package/dist/data/benchmarks.json

@@ -0,0 +1,61 @@
+{
+    "version": 1,
+    "entries": [
+        {
+            "modelId": "meta-llama/Llama-3-8B-Instruct",
+            "hardware": "gpu-low",
+            "avgLatencyMs": 850,
+            "measuredAt": "2026-08-01",
+            "promptTokens": 512,
+            "outputTokens": 256
+        },
+        {
+            "modelId": "meta-llama/Llama-3-8B-Instruct",
+            "hardware": "gpu-high",
+            "avgLatencyMs": 320,
+            "measuredAt": "2026-08-01",
+            "promptTokens": 512,
+            "outputTokens": 256
+        },
+        {
+            "modelId": "meta-llama/Llama-3-8B-Instruct",
+            "hardware": "cpu",
+            "avgLatencyMs": 9800,
+            "measuredAt": "2026-08-01",
+            "promptTokens": 512,
+            "outputTokens": 256
+        },
+        {
+            "modelId": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+            "hardware": "cpu",
+            "avgLatencyMs": 1400,
+            "measuredAt": "2026-08-01",
+            "promptTokens": 512,
+            "outputTokens": 256
+        },
+        {
+            "modelId": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+            "hardware": "gpu-low",
+            "avgLatencyMs": 180,
+            "measuredAt": "2026-08-01",
+            "promptTokens": 512,
+            "outputTokens": 256
+        },
+        {
+            "modelId": "mistralai/Mistral-7B-Instruct-v0.3",
+            "hardware": "gpu-low",
+            "avgLatencyMs": 790,
+            "measuredAt": "2026-08-01",
+            "promptTokens": 512,
+            "outputTokens": 256
+        },
+        {
+            "modelId": "mistralai/Mistral-7B-Instruct-v0.3",
+            "hardware": "gpu-high",
+            "avgLatencyMs": 290,
+            "measuredAt": "2026-08-01",
+            "promptTokens": 512,
+            "outputTokens": 256
+        }
+    ]
+}

package/dist/embeddings.d.ts

@@ -0,0 +1,50 @@
+/**
+ * A small, local, dependency-free text embedding scheme.
+ *
+ * WHY NOT A REAL NEURAL EMBEDDING MODEL: a genuine pretrained sentence
+ * embedding model (e.g. all-MiniLM-L6-v2 via transformers.js) needs to
+ * download model weights from HuggingFace at install/first-run time. This
+ * package's build/test environment can't verify that network path, so
+ * rather than ship an unverified dependency, this is a self-authored
+ * character-trigram hashing vectorizer + cosine similarity — a classic,
+ * well-understood, fully offline technique with zero external dependencies
+ * or license concerns (original code).
+ *
+ * TRADE-OFF: this captures lexical/surface similarity (shared substrings),
+ * not deep semantic meaning. "extract key-value pairs from invoices" and
+ * "structured data extraction" would score lower here than they would with
+ * a real neural embedding, since they share few character sequences despite
+ * being semantically close. Swap in a real model later by replacing embed()
+ * with one backed by transformers.js (or similar) — cosineSimilarity() and
+ * every caller of embed() would keep working unchanged, since only the
+ * vector-production step would change.
+ *
+ * SWAP-IN NOTE (build order step 6, "benchmark embedding-mode latency —
+ * feed into v0.2's benchmark dataset if relevant"): that benchmark dataset
+ * (src/data/benchmarks.json) measures LLM *inference* latency per
+ * (model, hardware) — a real, meaningful cost that varies by model size and
+ * hardware. This embedding scheme is a synchronous, local, sub-millisecond
+ * hash computation with no model inference involved, so there's no
+ * per-(model, hardware) latency to record there — "not relevant" for this
+ * implementation. See test/embeddings.test.ts's performance sanity check
+ * for a concrete timing figure instead. If embed() is ever swapped for a
+ * real neural embedding model (see the trade-off note above), *that* would
+ * introduce real inference latency worth feeding into the benchmark dataset.
+ */
+/**
+ * Produces a fixed-length (VECTOR_DIM), L2-normalized vector for the given
+ * text using hashed character-trigram counts. Deterministic: same text
+ * always produces the same vector.
+ */
+export declare function embed(text: string): number[];
+/**
+ * Cosine similarity between two equal-length vectors, in [0, 1] for these
+ * non-negative count-based vectors (in [-1, 1] in general). Since embed()
+ * always L2-normalizes, this is just the dot product.
+ */
+export declare function cosineSimilarity(a: number[], b: number[]): number;
+/** Exposed for tests. */
+export declare function clearEmbeddingCache(): void;
+/** Exposed for tests/debugging. */
+export declare function embeddingCacheSize(): number;
+//# sourceMappingURL=embeddings.d.ts.map

package/dist/embeddings.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"embeddings.d.ts","sourceRoot":"","sources":["../src/embeddings.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;AAwCH;;;;GAIG;AACH,wBAAgB,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,EAAE,CAmB5C;AAED;;;;GAIG;AACH,wBAAgB,gBAAgB,CAAC,CAAC,EAAE,MAAM,EAAE,EAAE,CAAC,EAAE,MAAM,EAAE,GAAG,MAAM,CAOjE;AAED,yBAAyB;AACzB,wBAAgB,mBAAmB,IAAI,IAAI,CAE1C;AAED,mCAAmC;AACnC,wBAAgB,kBAAkB,IAAI,MAAM,CAE3C"}

package/dist/embeddings.js

@@ -0,0 +1,118 @@
+"use strict";
+/**
+ * A small, local, dependency-free text embedding scheme.
+ *
+ * WHY NOT A REAL NEURAL EMBEDDING MODEL: a genuine pretrained sentence
+ * embedding model (e.g. all-MiniLM-L6-v2 via transformers.js) needs to
+ * download model weights from HuggingFace at install/first-run time. This
+ * package's build/test environment can't verify that network path, so
+ * rather than ship an unverified dependency, this is a self-authored
+ * character-trigram hashing vectorizer + cosine similarity — a classic,
+ * well-understood, fully offline technique with zero external dependencies
+ * or license concerns (original code).
+ *
+ * TRADE-OFF: this captures lexical/surface similarity (shared substrings),
+ * not deep semantic meaning. "extract key-value pairs from invoices" and
+ * "structured data extraction" would score lower here than they would with
+ * a real neural embedding, since they share few character sequences despite
+ * being semantically close. Swap in a real model later by replacing embed()
+ * with one backed by transformers.js (or similar) — cosineSimilarity() and
+ * every caller of embed() would keep working unchanged, since only the
+ * vector-production step would change.
+ *
+ * SWAP-IN NOTE (build order step 6, "benchmark embedding-mode latency —
+ * feed into v0.2's benchmark dataset if relevant"): that benchmark dataset
+ * (src/data/benchmarks.json) measures LLM *inference* latency per
+ * (model, hardware) — a real, meaningful cost that varies by model size and
+ * hardware. This embedding scheme is a synchronous, local, sub-millisecond
+ * hash computation with no model inference involved, so there's no
+ * per-(model, hardware) latency to record there — "not relevant" for this
+ * implementation. See test/embeddings.test.ts's performance sanity check
+ * for a concrete timing figure instead. If embed() is ever swapped for a
+ * real neural embedding model (see the trade-off note above), *that* would
+ * introduce real inference latency worth feeding into the benchmark dataset.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.embed = embed;
+exports.cosineSimilarity = cosineSimilarity;
+exports.clearEmbeddingCache = clearEmbeddingCache;
+exports.embeddingCacheSize = embeddingCacheSize;
+const VECTOR_DIM = 256;
+const MAX_CACHE_SIZE = 500;
+/** djb2 string hash — fast, deterministic, decent distribution for this use. */
+function hashString(s) {
+    let hash = 5381;
+    for (let i = 0; i < s.length; i++) {
+        hash = ((hash << 5) + hash + s.charCodeAt(i)) >>> 0;
+    }
+    return hash;
+}
+function normalizeText(text) {
+    return text
+        .toLowerCase()
+        .replace(/[^a-z0-9\s]/g, " ")
+        .replace(/\s+/g, " ")
+        .trim();
+}
+function charTrigrams(text) {
+    const padded = ` ${normalizeText(text)} `;
+    const grams = [];
+    for (let i = 0; i < padded.length - 2; i++) {
+        grams.push(padded.slice(i, i + 3));
+    }
+    return grams;
+}
+/**
+ * Cache of computed embeddings, keyed by exact input text. Per the v0.3
+ * Do's/Don'ts ("don't recompute embeddings for the same model description
+ * on every call"), repeated calls with identical text are served from here
+ * instead of re-hashing. Simple FIFO eviction once MAX_CACHE_SIZE is hit —
+ * this is a small in-process cache, not meant to be a durable store.
+ */
+const embeddingCache = new Map();
+/**
+ * Produces a fixed-length (VECTOR_DIM), L2-normalized vector for the given
+ * text using hashed character-trigram counts. Deterministic: same text
+ * always produces the same vector.
+ */
+function embed(text) {
+    const cached = embeddingCache.get(text);
+    if (cached)
+        return cached;
+    const vector = new Array(VECTOR_DIM).fill(0);
+    for (const gram of charTrigrams(text)) {
+        vector[hashString(gram) % VECTOR_DIM] += 1;
+    }
+    const norm = Math.sqrt(vector.reduce((sum, v) => sum + v * v, 0));
+    const normalized = norm > 0 ? vector.map((v) => v / norm) : vector;
+    if (embeddingCache.size >= MAX_CACHE_SIZE) {
+        const oldestKey = embeddingCache.keys().next().value;
+        if (oldestKey !== undefined)
+            embeddingCache.delete(oldestKey);
+    }
+    embeddingCache.set(text, normalized);
+    return normalized;
+}
+/**
+ * Cosine similarity between two equal-length vectors, in [0, 1] for these
+ * non-negative count-based vectors (in [-1, 1] in general). Since embed()
+ * always L2-normalizes, this is just the dot product.
+ */
+function cosineSimilarity(a, b) {
+    if (a.length !== b.length) {
+        throw new Error(`Vectors must be the same length (got ${a.length} and ${b.length})`);
+    }
+    let dot = 0;
+    for (let i = 0; i < a.length; i++)
+        dot += a[i] * b[i];
+    return dot;
+}
+/** Exposed for tests. */
+function clearEmbeddingCache() {
+    embeddingCache.clear();
+}
+/** Exposed for tests/debugging. */
+function embeddingCacheSize() {
+    return embeddingCache.size;
+}
+//# sourceMappingURL=embeddings.js.map

package/dist/embeddings.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"embeddings.js","sourceRoot":"","sources":["../src/embeddings.ts"],"names":[],"mappings":";AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;;;;;;AAEH,MAAM,UAAU,GAAG,GAAG,CAAC;AACvB,MAAM,cAAc,GAAG,GAAG,CAAC;AAE3B,gFAAgF;AAChF,SAAS,UAAU,CAAC,CAAS;IAC3B,IAAI,IAAI,GAAG,IAAI,CAAC;IAChB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAClC,IAAI,GAAG,CAAC,CAAC,IAAI,IAAI,CAAC,CAAC,GAAG,IAAI,GAAG,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;IACtD,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC;AAED,SAAS,aAAa,CAAC,IAAY;IACjC,OAAO,IAAI;SACR,WAAW,EAAE;SACb,OAAO,CAAC,cAAc,EAAE,GAAG,CAAC;SAC5B,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC;SACpB,IAAI,EAAE,CAAC;AACZ,CAAC;AAED,SAAS,YAAY,CAAC,IAAY;IAChC,MAAM,MAAM,GAAG,IAAI,aAAa,CAAC,IAAI,CAAC,GAAG,CAAC;IAC1C,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC3C,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;IACrC,CAAC;IACD,OAAO,KAAK,CAAC;AACf,CAAC;AAED;;;;;;GAMG;AACH,MAAM,cAAc,GAAG,IAAI,GAAG,EAAoB,CAAC;AAEnD;;;;GAIG;AACH,eAAsB,IAAY;IAChC,MAAM,MAAM,GAAG,cAAc,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;IACxC,IAAI,MAAM;QAAE,OAAO,MAAM,CAAC;IAE1B,MAAM,MAAM,GAAG,IAAI,KAAK,CAAC,UAAU,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAC7C,KAAK,MAAM,IAAI,IAAI,YAAY,CAAC,IAAI,CAAC,EAAE,CAAC;QACtC,MAAM,CAAC,UAAU,CAAC,IAAI,CAAC,GAAG,UAAU,CAAC,IAAI,CAAC,CAAC;IAC7C,CAAC;IAED,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;IAClE,MAAM,UAAU,GAAG,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC;IAEnE,IAAI,cAAc,CAAC,IAAI,IAAI,cAAc,EAAE,CAAC;QAC1C,MAAM,SAAS,GAAG,cAAc,CAAC,IAAI,EAAE,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC;QACrD,IAAI,SAAS,KAAK,SAAS;YAAE,cAAc,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC;IAChE,CAAC;IACD,cAAc,CAAC,GAAG,CAAC,IAAI,EAAE,UAAU,CAAC,CAAC;IAErC,OAAO,UAAU,CAAC;AACpB,CAAC;AAED;;;;GAIG;AACH,0BAAiC,CAAW,EAAE,CAAW;IACvD,IAAI,CAAC,CAAC,MAAM,KAAK,CAAC,CAAC,MAAM,EAAE,CAAC;QAC1B,MAAM,IAAI,KAAK,CAAC,wCAAwC,CAAC,CAAC,MAAM,QAAQ,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC;IACvF,CAAC;IACD,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE;QAAE,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;IACtD,OAAO,GAAG,CAAC;AACb,CAAC;AAED,yBAAyB;AACzB;IACE,cAAc,CAAC,KAAK,EAAE,CAAC;AACzB,CAAC;AAED,mCAAmC;AACnC;IACE,OAAO,cAAc,CAAC,IAAI,CAAC;AAC7B,CAAC"}

package/dist/errors.d.ts

@@ -0,0 +1,18 @@
+/** Base class for every error smallm throws. Lets callers do `instanceof SmallmError`. */
+export declare class SmallmError extends Error {
+    constructor(message: string);
+}
+/** Non-2xx response from the HuggingFace API (excluding the 429 case, see RateLimitError). */
+export declare class HFApiError extends SmallmError {
+    readonly status?: number | undefined;
+    constructor(message: string, status?: number | undefined);
+}
+/** Specifically a 429 from HuggingFace. Triggers retry/backoff — see registry/huggingface.ts. */
+export declare class RateLimitError extends HFApiError {
+    constructor(message: string, status?: number);
+}
+/** Bad ModelQuery input. Never retried — a malformed request will never succeed by retrying. */
+export declare class ValidationError extends SmallmError {
+    constructor(message: string);
+}
+//# sourceMappingURL=errors.d.ts.map

package/dist/errors.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"errors.d.ts","sourceRoot":"","sources":["../src/errors.ts"],"names":[],"mappings":"AAAA,0FAA0F;AAC1F,qBAAa,WAAY,SAAQ,KAAK;IACpC,YAAY,OAAO,EAAE,MAAM,EAK1B;CACF;AAED,8FAA8F;AAC9F,qBAAa,UAAW,SAAQ,WAAW;aACI,MAAM,CAAC,EAAE,MAAM;IAA5D,YAAY,OAAO,EAAE,MAAM,EAAkB,MAAM,CAAC,EAAE,MAAM,YAAA,EAI3D;CACF;AAED,iGAAiG;AACjG,qBAAa,cAAe,SAAQ,UAAU;IAC5C,YAAY,OAAO,EAAE,MAAM,EAAE,MAAM,GAAE,MAAY,EAIhD;CACF;AAED,gGAAgG;AAChG,qBAAa,eAAgB,SAAQ,WAAW;IAC9C,YAAY,OAAO,EAAE,MAAM,EAI1B;CACF"}

package/dist/errors.js

@@ -0,0 +1,42 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.ValidationError = exports.RateLimitError = exports.HFApiError = exports.SmallmError = void 0;
+/** Base class for every error smallm throws. Lets callers do `instanceof SmallmError`. */
+class SmallmError extends Error {
+    constructor(message) {
+        super(message);
+        this.name = "SmallmError";
+        // Restore prototype chain (needed when compiling to older targets / extending Error in TS).
+        Object.setPrototypeOf(this, new.target.prototype);
+    }
+}
+exports.SmallmError = SmallmError;
+/** Non-2xx response from the HuggingFace API (excluding the 429 case, see RateLimitError). */
+class HFApiError extends SmallmError {
+    constructor(message, status) {
+        super(message);
+        this.status = status;
+        this.name = "HFApiError";
+        Object.setPrototypeOf(this, new.target.prototype);
+    }
+}
+exports.HFApiError = HFApiError;
+/** Specifically a 429 from HuggingFace. Triggers retry/backoff — see registry/huggingface.ts. */
+class RateLimitError extends HFApiError {
+    constructor(message, status = 429) {
+        super(message, status);
+        this.name = "RateLimitError";
+        Object.setPrototypeOf(this, new.target.prototype);
+    }
+}
+exports.RateLimitError = RateLimitError;
+/** Bad ModelQuery input. Never retried — a malformed request will never succeed by retrying. */
+class ValidationError extends SmallmError {
+    constructor(message) {
+        super(message);
+        this.name = "ValidationError";
+        Object.setPrototypeOf(this, new.target.prototype);
+    }
+}
+exports.ValidationError = ValidationError;
+//# sourceMappingURL=errors.js.map

package/dist/errors.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"errors.js","sourceRoot":"","sources":["../src/errors.ts"],"names":[],"mappings":";;;AAAA,0FAA0F;AAC1F,iBAAyB,SAAQ,KAAK;IACpC,YAAY,OAAe;QACzB,KAAK,CAAC,OAAO,CAAC,CAAC;QACf,IAAI,CAAC,IAAI,GAAG,aAAa,CAAC;QAC1B,4FAA4F;QAC5F,MAAM,CAAC,cAAc,CAAC,IAAI,EAAE,IAAI,MAAM,CAAC,SAAS,CAAC,CAAC;IACpD,CAAC;CACF;;AAED,8FAA8F;AAC9F,gBAAwB,SAAQ,WAAW;IACzC,YAAY,OAAe,EAAkB,MAAe;QAC1D,KAAK,CAAC,OAAO,CAAC,CAAC;sBAD4B,MAAM;QAEjD,IAAI,CAAC,IAAI,GAAG,YAAY,CAAC;QACzB,MAAM,CAAC,cAAc,CAAC,IAAI,EAAE,IAAI,MAAM,CAAC,SAAS,CAAC,CAAC;IACpD,CAAC;CACF;;AAED,iGAAiG;AACjG,oBAA4B,SAAQ,UAAU;IAC5C,YAAY,OAAe,EAAE,MAAM,GAAW,GAAG;QAC/C,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC;QACvB,IAAI,CAAC,IAAI,GAAG,gBAAgB,CAAC;QAC7B,MAAM,CAAC,cAAc,CAAC,IAAI,EAAE,IAAI,MAAM,CAAC,SAAS,CAAC,CAAC;IACpD,CAAC;CACF;;AAED,gGAAgG;AAChG,qBAA6B,SAAQ,WAAW;IAC9C,YAAY,OAAe;QACzB,KAAK,CAAC,OAAO,CAAC,CAAC;QACf,IAAI,CAAC,IAAI,GAAG,iBAAiB,CAAC;QAC9B,MAAM,CAAC,cAAc,CAAC,IAAI,EAAE,IAAI,MAAM,CAAC,SAAS,CAAC,CAAC;IACpD,CAAC;CACF"}

package/dist/filters.d.ts

@@ -0,0 +1,11 @@
+import { HFModel, ModelQuery } from "./types";
+/**
+ * Applies hard filters before any scoring happens. A model failing ANY
+ * filter is excluded outright — it never reaches scorer.ts. Unknown values
+ * never cause exclusion (MVP Section 4.4's rule, carried forward per the
+ * v0.2/v0.3/v0.4 guides' "unknown data is never punished" cross-phase
+ * principle). Provider-agnostic throughout — never branches on
+ * model.provider (v0.4 Do's/Don'ts).
+ */
+export declare function applyHardFilters(models: HFModel[], query: ModelQuery): HFModel[];
+//# sourceMappingURL=filters.d.ts.map

package/dist/filters.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"filters.d.ts","sourceRoot":"","sources":["../src/filters.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,MAAM,SAAS,CAAC;AAe9C;;;;;;;GAOG;AACH,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,OAAO,EAAE,EAAE,KAAK,EAAE,UAAU,GAAG,OAAO,EAAE,CAiDhF"}