@theglitchking/semantic-memory 1.6.0 → 1.7.0

package/hooks/jev-bridge.py

@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+"""semantic-memory → Jev, through jev-loop's router (v1.7 jev mode). Stdlib only.
+
+    echo '{"cwd": "...", "caller": "semantic-memory-judge", "state": "...",
+           "questions": [{"key": "a", "instructions": "..."}]}' | python3 jev-bridge.py
+
+Prints one JSON line: {"scores": {key: probability}, "tokens": n} or {"scores": {}, "error": "..."}.
+Never raises, never exits non-zero: every failure is an empty answer with a reason.
+
+The only way semantic-memory reaches TypeSafe is `router.ask`, which redacts the state and every
+question (redact.py), caches answers, writes the usage ledger under the caller's name and enforces the
+batch budget. This file never builds a request itself and never prints the key.
+
+Callers: semantic-memory-hook (the failure hook) and semantic-memory-judge / semantic-memory-skill-loop.
+jev-loop counts all three as batch (capped and cached) until it lists semantic-memory-hook as live. The
+hook caller has its own daily ceiling here, SEMANTIC_MEMORY_JEV_DAILY_TOKENS (default 500000).
+"""
+import json
+import os
+import sys
+from datetime import datetime
+from pathlib import Path
+
+CALLERS = {"semantic-memory-hook", "semantic-memory-judge", "semantic-memory-skill-loop"}
+LIVE_CALLER = "semantic-memory-hook"
+
+
+def out(obj):
+    sys.stdout.write(json.dumps(obj) + "\n")
+    sys.stdout.flush()
+
+
+def live_tokens_today(logs, caller):
+    """Tokens this caller spent today, from jev-loop's usage ledger (numbers only)."""
+    day, used = datetime.now().date().isoformat(), 0
+    for path in Path(logs).glob("jev-usage-*.jsonl"):
+        for line in path.read_text().splitlines():
+            try:
+                r = json.loads(line)
+            except ValueError:
+                continue
+            if r.get("caller") == caller and str(r.get("time", "")).startswith(day):
+                used += r.get("input_tokens", 0) + r.get("output_tokens", 0)
+    return used
+
+
+def main():
+    try:
+        req = json.loads(sys.stdin.read() or "{}")
+    except ValueError as e:
+        return out({"scores": {}, "error": f"bad input: {e}"})
+    caller = req.get("caller")
+    if caller not in CALLERS:
+        return out({"scores": {}, "error": f"unknown caller {caller!r}"})
+    state, questions = req.get("state"), req.get("questions") or []
+    if not isinstance(state, str) or not state or not questions:
+        return out({"scores": {}, "error": "state and questions are required"})
+    if not all(isinstance(q, dict) and q.get("key") and isinstance(q.get("instructions"), str) for q in questions):
+        return out({"scores": {}, "error": "each question needs a key and instructions"})
+
+    jev_dir = Path(os.environ.get("JEV_LOOP_DIR") or Path.home() / ".claude" / "jev-loop")
+    if not (jev_dir / "router.py").exists():
+        return out({"scores": {}, "error": f"jev-loop not found at {jev_dir}"})
+    if not os.environ.get("TYPESAFE_API_KEY"):
+        # cron and plain shells have no Claude settings env: read the key the way jev-loop's daily.sh does
+        try:
+            settings = Path(os.environ.get("SEMANTIC_MEMORY_CLAUDE_SETTINGS") or Path.home() / ".claude" / "settings.json")
+            os.environ["TYPESAFE_API_KEY"] = json.loads(settings.read_text())["env"]["TYPESAFE_API_KEY"]
+        except Exception:
+            return out({"scores": {}, "error": "TYPESAFE_API_KEY is not set"})
+
+    # the router names its caller from argv[0] (older jev-loop) or JEV_CALLER: set both before importing it
+    os.environ["JEV_CALLER"] = caller
+    sys.argv = [caller]
+    sys.path.insert(0, str(jev_dir))
+    try:
+        import router  # noqa: E402
+    except Exception as e:
+        return out({"scores": {}, "error": f"router import failed: {type(e).__name__}: {e}"})
+
+    cwd = req.get("cwd") or os.getcwd()
+    if router.repo_for(cwd, []) == "excluded":
+        return out({"scores": {}, "error": "excluded path: nothing sent"})
+
+    if caller == LIVE_CALLER:
+        cap = int(os.environ.get("SEMANTIC_MEMORY_JEV_DAILY_TOKENS") or 500_000)
+        used = live_tokens_today(router.LOGS, caller)
+        if used >= cap:
+            return out({"scores": {}, "error": f"daily ceiling reached: {used} of {cap} tokens (SEMANTIC_MEMORY_JEV_DAILY_TOKENS)", "ceiling": True})
+    timeout = req.get("timeout_seconds")
+    if isinstance(timeout, (int, float)) and timeout > 0:
+        router.R["timeout_seconds"] = timeout
+
+    state = state[: router.R["max_prompt_chars"]]
+    items = [(q["key"], q["instructions"]) for q in questions]
+    try:
+        scores, tokens = router.ask_all(state, items)
+    except Exception as e:  # JevBudget, HTTP errors, timeouts: no answer, never a crash
+        router.flush_ledger()
+        return out({"scores": {}, "error": f"{type(e).__name__}: {e}"[:300]})
+    router.flush_ledger()
+    out({"scores": {k: v for k, v in scores.items() if isinstance(v, (int, float))}, "tokens": tokens})
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except BaseException as e:  # noqa: BLE001 — the bridge must never crash its caller
+        out({"scores": {}, "error": f"{type(e).__name__}: {e}"[:300]})

package/hooks/session-start.js

@@ -18,6 +18,11 @@ import { reconcile } from "./reconcile.js";
 
 async function main() {
   const projectRoot = process.env.CLAUDE_PROJECT_DIR || process.cwd();
+  // recursion guard: a skill-loop claude run (v1.7) must not reconcile or update anything
+  if (process.env.SEMANTIC_MEMORY_REVIEWER === "1") {
+    emit();
+    return;
+  }
 
   // 1. Reconcile — no-deps, always runs.
   try {

package/hooks/skill-loop.js

@@ -0,0 +1,65 @@
+#!/usr/bin/env node
+// skill-loop.js — the skill loop's hooks (v1.7). Off by default: with the skill loop off (or inside a
+// skill-loop claude run) every event leaves at once, writes nothing and spawns nothing.
+//
+//   SessionStart                → once per session (startup / resume): "N skill drafts waiting"; starts the curator
+//                                 detached when it is due (every 7 days, repo quiet >= 2 h)
+//   PostToolUse (matcher Skill) → one usage line in .claude/.semantic-memory/skill-usage.jsonl
+//   SessionEnd                  → a review job when the session has signal (a failure, a correction or many
+//                                 tool calls), and the repo's worker started detached. Must finish well inside
+//                                 SessionEnd's 1.5 s: it only writes the job and spawns.
+//
+// Never blocks the session and never fails it: errors are swallowed (SIDEKICK_DEBUG=1 prints them).
+
+import { readFileSync } from "node:fs";
+import { resolve } from "node:path";
+import { fileURLToPath } from "node:url";
+
+const DEBUG = process.env.SIDEKICK_DEBUG === "1";
+const debug = (msg) => DEBUG && process.stderr.write(`[skill-loop] ${msg}\n`);
+const dist = (m) => new URL(`../dist/core/${m}.js`, import.meta.url).href;
+
+function readStdinJson() {
+  try {
+    const raw = readFileSync(0, "utf8");
+    return raw.trim() ? JSON.parse(raw) : {};
+  } catch {
+    return {};
+  }
+}
+
+export async function main(input = readStdinJson(), env = process.env) {
+  if (env.SEMANTIC_MEMORY_REVIEWER === "1") return;
+  const cwd = input.cwd || env.CLAUDE_PROJECT_DIR || process.cwd();
+  const { featureOn } = await import(dist("features"));
+  if (!featureOn("skill_loop", cwd, env)) return;
+  const ev = await import(dist("skill-evidence"));
+  const root = ev.repoRoot(cwd);
+  const session_id = typeof input.session_id === "string" ? input.session_id : null;
+  if (input.hook_event_name === "SessionStart") {
+    if (input.source === "clear" || input.source === "compact") return;
+    const { listDrafts } = await import(dist("skill-review"));
+    const n = listDrafts(root).length;
+    const { curatorDue } = await import(dist("skill-curator"));
+    if (curatorDue(root)) {
+      const { spawn } = await import("node:child_process");
+      const cli = new URL("../dist/cli/index.js", import.meta.url).pathname;
+      spawn(process.execPath, [cli, "skill-loop", "curate", "--auto"], { cwd: root, env, detached: true, stdio: "ignore" }).unref();
+    }
+    if (n) process.stdout.write(JSON.stringify({ hookSpecificOutput: { hookEventName: "SessionStart", additionalContext: `[skill loop] ${n} skill draft${n > 1 ? "s" : ""} waiting for review (\`/skill-loop review\`).` } }) + "\n");
+  } else if (input.hook_event_name === "PostToolUse" && input.tool_name === "Skill") {
+    const skill = input.tool_input?.skill;
+    if (typeof skill === "string" && skill) ev.appendEvidence(root, "usage", { session_id, skill, via: "tool" }, env);
+  } else if (input.hook_event_name === "SessionEnd") {
+    const q = await import(dist("skill-queue"));
+    const job = q.sessionJob(root, input, env);
+    if (!job) return;
+    q.enqueue(job, env);
+    const { spawnWorker } = await import(dist("skill-worker"));
+    spawnWorker(root, env);
+  }
+}
+
+if (process.argv[1] && resolve(process.argv[1]) === fileURLToPath(import.meta.url)) {
+  main().catch((e) => debug(e.message));
+}

package/hooks/vault-context.js

@@ -939,7 +939,32 @@ async function handleStop(projectRoot, sessionId) {
 // --- PostToolUseFailure: look up a knowledge-base fact for a failed tool call (v1.6, opt-in) ---
 // Measured on 553 real failures (scripts/eval-failures): facts-only at score >= 0.50, about 6 in 10
 // injections were on target, firing on ~4% of failures. Below the plan's 95% bar, hence off by default.
-const FAILURE_HOOK_ON = () => process.env.SEMANTIC_MEMORY_FAILURE_HOOK === "1";
+// v1.7 jev mode: Jev can veto the candidate fact. Measured on 1,171 Claude-labelled failure/doc pairs, Jev's
+// confident no's (<= 0.05) were 727 of 729 right, but it never said yes at >= 0.95 to a relevant pair (max 0.89):
+// a "yes" gate would inject nothing. So a confident no suppresses, anything else keeps the plain hook's choice.
+// Replayed, the veto removed only wrong facts (12 wrong, 0 right at score 0.30) but none at the shipped 0.50, and
+// precision stays far from the 95% bar, so jev mode does not turn the hook on by itself.
+const JEV_GATE = { adopted: false, budgetMs: Number(process.env.SEMANTIC_MEMORY_JEV_HOOK_BUDGET_MS) || 1500 };
+const FAILURE_HOOK_ENV = () => process.env.SEMANTIC_MEMORY_FAILURE_HOOK;
+const FAILURE_HOOK_ON = (projectRoot) =>
+  FAILURE_HOOK_ENV() === "1" || (FAILURE_HOOK_ENV() !== "0" && JEV_GATE.adopted && jevOn(projectRoot));
+
+// cheap (env + a few stats); the hook-side copy of src/core/features.ts
+let featuresMod;
+async function loadFeatures() {
+  if (featuresMod === undefined) {
+    try {
+      featuresMod = await import(new URL("../dist/core/features.js", import.meta.url).href);
+    } catch (e) {
+      debug(`features unavailable: ${e.message}`);
+      featuresMod = null;
+    }
+  }
+  return featuresMod;
+}
+function jevOn(projectRoot) {
+  return Boolean(featuresMod && featuresMod.featureOn("jev", projectRoot));
+}
 const FAILURE_MIN_SCORE = Number(process.env.SEMANTIC_MEMORY_FAILURE_MIN_SCORE) || 0.5;
 const USER_DENIED = /^(The user doesn't want to proceed|Permission to use .* has been denied)/;
 const isFactPath = (p) => /(^|\/)facts\/[^/]+\.md$/.test(String(p));
@@ -964,7 +989,65 @@ async function redactText(text) {
   }
 }
 
-async function handleFailure(projectRoot, sources, cliBin, input, sessionId) {
+// One live Jev call for the candidate: a confident no (<= 0.05) vetoes it, a yes or an unsure answer lets it
+// through. An error, a timeout or the daily ceiling injects nothing: in jev mode an unchecked fact is not shown.
+async function jevAllows(projectRoot, command, error, fact, started) {
+  try {
+    const jev = await import(new URL("../dist/core/jev.js", import.meta.url).href);
+    const title = String(fact.title ?? fact.path);
+    const excerpt = await redactText(String(fact.snippet ?? ""));
+    const left = JEV_GATE.budgetMs - (Date.now() - started);
+    if (left < 300) {
+      debug(`jev gate: ${left} ms left of the hook budget — skipped`);
+      return false;
+    }
+    const a = await jev.askJev(
+      {
+        cwd: projectRoot,
+        caller: "semantic-memory-hook",
+        state: jev.failureState(command, error),
+        questions: [{ key: "fits", instructions: jev.relevanceQuestion(title, excerpt) }],
+      },
+      { timeoutMs: left - 50 },
+    );
+    if (a.ceiling) process.stderr.write(`[semantic-memory] jev mode: ${a.error}; the failure hook injects nothing until tomorrow\n`);
+    else if (a.error) debug(`jev gate: ${a.error}`);
+    if (a.error || typeof a.scores.fits !== "number") return false;
+    return jev.jevVerdict(a.scores.fits) !== false;
+  } catch (e) {
+    debug(`jev gate failed: ${e.message}`);
+    return false;
+  }
+}
+
+// v1.7 skill loop: failures, corrections and /skill uses become evidence for the between-session review.
+// Never blocks or changes this hook's output; any error is swallowed.
+async function recordSkillEvidence(eventName, projectRoot, input) {
+  try {
+    const ev = await import(new URL("../dist/core/skill-evidence.js", import.meta.url).href);
+    const root = ev.repoRoot(projectRoot);
+    const session_id = typeof input.session_id === "string" ? input.session_id : null;
+    if (eventName === "PostToolUseFailure") {
+      const name = String(input.tool_name || "");
+      const error = String(input.error || "");
+      if ((name === "Bash" || name.startsWith("mcp__")) && !input.is_interrupt && error && !USER_DENIED.test(error.trim())) {
+        ev.appendEvidence(root, "failures", ev.failureRecord(input));
+      }
+    } else if (eventName === "UserPromptSubmit" && typeof input.prompt === "string") {
+      const cue = ev.detectCorrection(input.prompt);
+      if (cue) ev.appendEvidence(root, "corrections", { session_id, cue, excerpt: input.prompt.slice(0, 300) });
+      const slash = ev.slashSkill(input.prompt);
+      if (slash) {
+        const file = join(root, ".claude", "skills", slash, "SKILL.md");
+        if (existsSync(file) && /^skill_loop:/m.test(readFileSync(file, "utf8"))) ev.appendEvidence(root, "usage", { session_id, skill: slash, via: "slash" });
+      }
+    }
+  } catch (e) {
+    debug(`skill evidence failed: ${e.message}`);
+  }
+}
+
+async function handleFailure(projectRoot, sources, cliBin, input, sessionId, started = Date.now()) {
   const name = String(input.tool_name || "");
   const error = String(input.error || "");
   if (!(name === "Bash" || name.startsWith("mcp__")) || input.is_interrupt || !error || USER_DENIED.test(error.trim())) {
@@ -979,11 +1062,16 @@ async function handleFailure(projectRoot, sources, cliBin, input, sessionId) {
   const command = typeof args.command === "string" ? args.command : `${name} ${JSON.stringify(args)}`.slice(0, 300);
   const query = await failureQueryFor(error, command);
   const hits = await searchSources(cliBin, sources, 8, query);
+  const gated = jevOn(projectRoot);
   const fact = hits.find((h) => isFactPath(h.path) && typeof h.score === "number" && h.score >= FAILURE_MIN_SCORE);
   if (!fact) {
     emit("PostToolUseFailure", "");
     return;
   }
+  if (gated && !(await jevAllows(projectRoot, command, error, fact, started))) {
+    emit("PostToolUseFailure", "");
+    return;
+  }
 
   // the same fact at most once per session: a repeat failure the agent already saw the fact for adds only noise
   const { resolvedRead, writePath } = statePath(projectRoot, "failure-shown");
@@ -1006,8 +1094,10 @@ async function handleFailure(projectRoot, sources, cliBin, input, sessionId) {
   }
 
   const lines = [
-    `<vault-fact source="tool-failure" score="${fact.score.toFixed(2)}">`,
-    `A knowledge-base fact may explain this failure. Check it applies before acting on it (measured: about 6 in 10 of these are on target).`,
+    `<vault-fact source="tool-failure" score="${fact.score.toFixed(2)}"${gated ? ` checked="jev"` : ""}>`,
+    gated
+      ? `A knowledge-base fact may explain this failure (checked: not ruled out). Check it applies before acting on it.`
+      : `A knowledge-base fact may explain this failure. Check it applies before acting on it (measured: about 6 in 10 of these are on target).`,
     `- \`${fact.path}\`${fact.source ? ` (${fact.source})` : ""}`,
   ];
   if (fact.snippet) lines.push(`  > ${(await redactText(String(fact.snippet))).replace(/\s+/g, " ").slice(0, 240)}`);
@@ -1066,8 +1156,16 @@ async function main() {
   const projectRoot = input.cwd || process.env.CLAUDE_PROJECT_DIR || process.cwd();
   const sessionId = typeof input.session_id === "string" ? input.session_id : undefined;
   debug(`event=${eventName} projectRoot=${projectRoot}`);
+  const started = Date.now();
+  // recursion guard: inside a skill-loop claude run every semantic-memory hook leaves at once
+  if (process.env.SEMANTIC_MEMORY_REVIEWER === "1") {
+    emitNoop(eventName);
+    return;
+  }
+  await loadFeatures();
+  if (featuresMod?.featureOn("skill_loop", projectRoot)) await recordSkillEvidence(eventName, projectRoot, input);
   // opt-in, and registered for every failed tool call: leave before any lookup when off
-  if (eventName === "PostToolUseFailure" && !FAILURE_HOOK_ON()) {
+  if (eventName === "PostToolUseFailure" && !FAILURE_HOOK_ON(projectRoot)) {
     emitNoop(eventName);
     return;
   }
@@ -1091,7 +1189,7 @@ async function main() {
     if (eventName === "UserPromptSubmit") {
       await handlePrompt(projectRoot, sources, cliBin, input);
     } else if (eventName === "PostToolUseFailure") {
-      await handleFailure(projectRoot, sources, cliBin, input, sessionId);
+      await handleFailure(projectRoot, sources, cliBin, input, sessionId, started);
     } else if (eventName === "Stop") {
       await handleStop(projectRoot, sessionId);
     } else {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@theglitchking/semantic-memory",
-  "version": "1.6.0",
+  "version": "1.7.0",
   "description": "Unified memory-layer plugin for Claude Code. One MCP server, multiple typed corpora (vault, code, plans, docs, research, project-map), unified cross-corpus knowledge graph, conditional tool registration, in-process drift detection, and human↔code translation. Rebrand of @theglitchking/semantic-sidekick at v1.0 — see CHANGELOG for migration.",
   "type": "module",
   "main": "./dist/core/index.js",
@@ -29,7 +29,9 @@
     "README.md",
     "CHANGELOG.md",
     "ROADMAP.md",
-    "LICENSE"
+    "LICENSE",
+    "THIRD_PARTY_NOTICES.md",
+    "LICENSES"
   ],
   "scripts": {
     "build": "tsup",

package/dist/chunk-44BQDRXD.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/core/embedder.ts","../src/core/graph.ts","../src/core/vector.ts","../src/core/search-text.ts"],"sourcesContent":["import { AutoTokenizer } from \"@huggingface/transformers\";\nimport { readFile, writeFile, mkdir, rename, unlink } from \"node:fs/promises\";\nimport { join, dirname } from \"node:path\";\nimport { existsSync, createWriteStream } from \"node:fs\";\nimport { homedir } from \"node:os\";\nimport { Readable } from \"node:stream\";\nimport { pipeline as streamPipeline } from \"node:stream/promises\";\nimport { Worker } from \"node:worker_threads\";\nimport { fileURLToPath } from \"node:url\";\n\n// MiniLM-L6-v2 is the default: ~3 min to index 2,853 chunks on CPU vs ~16 min for nomic.\n// nomic-embed-text-v1.5 gives higher quality embeddings but is much slower on CPU.\nconst DEFAULT_MODEL = \"sentence-transformers/all-MiniLM-L6-v2\";\nconst CACHE_DIR = join(homedir(), \".semantic-sidekick\", \"models\");\n// Default to 1 worker (serial). Worker threads only help on memory-rich machines\n// (each worker loads its own ONNX session). On typical dev machines\n// with <4 GB free RAM, parallel workers cause swap thrashing and are slower.\n// Enable with --workers N when you have sufficient RAM.\nconst DEFAULT_WORKERS = 1;\n// batch=16 is optimal for MiniLM on CPU (short sequences, low padding waste).\n// For larger models (nomic 768d), batch=1 is faster due to padding overhead.\nconst DEFAULT_BATCH_SIZE = 16;\n// Quantized ONNX: faster on CPU but not all models have a quantized variant.\n// MiniLM fp32 is already fast enough; nomic benefits from quantized.\n// Falls back to fp32 automatically if quantized file is not available.\nconst DEFAULT_QUANTIZED = false;\n\n// Full-precision ONNX model subpaths\nconst ONNX_MODEL_PATHS: Record<string, string> = {\n  \"nomic-ai/nomic-embed-text-v1.5\": \"onnx/model.onnx\",\n  \"sentence-transformers/all-MiniLM-L6-v2\": \"onnx/model.onnx\",\n};\n\n// Quantized (int8) ONNX model subpaths — faster on CPU, ~same quality\nconst ONNX_QUANTIZED_MODEL_PATHS: Record<string, string> = {\n  \"nomic-ai/nomic-embed-text-v1.5\": \"onnx/model_quantized.onnx\",\n  \"sentence-transformers/all-MiniLM-L6-v2\": \"onnx/model_quantized.onnx\",\n};\n\ninterface OrtSession {\n  run(feeds: Record<string, unknown>): Promise<Record<string, { data: Float32Array; dims: number[] }>>;\n  inputNames: string[];\n  outputNames: string[];\n}\n\ninterface OrtModule {\n  InferenceSession: {\n    create(path: string, options?: Record<string, unknown>): Promise<OrtSession>;\n  };\n  Tensor: new (type: string, data: ArrayLike<number | bigint>, dims: number[]) => unknown;\n}\n\ntype RuntimeLabel = \"native\" | \"wasm\";\n\nasync function resolveOnnxRuntime(): Promise<{ ort: OrtModule; label: RuntimeLabel }> {\n  try {\n    const ort = await import(\"onnxruntime-node\");\n    return { ort: ort as unknown as OrtModule, label: \"native\" };\n  } catch {\n    const ort = await import(\"onnxruntime-web\");\n    return { ort: ort as unknown as OrtModule, label: \"wasm\" };\n  }\n}\n\n// Download to a process-unique temp file, then atomically rename onto the\n// final path. Two concurrent downloaders won't corrupt each other's partial\n// writes — the worst case is one wasted download (rename is last-writer-wins\n// but each rename installs a complete, valid file). Without this, multiple\n// processes calling Embedder.init() against the same cache (e.g. parallel\n// vitest workers, or multiple semantic-sidekick servers on a fresh install) race\n// on a single shared writeStream and produce a corrupt ONNX file that fails\n// Protobuf parsing on load.\nasync function downloadFile(url: string, destPath: string): Promise<void> {\n  const response = await fetch(url);\n  if (!response.ok) throw new Error(`Download failed (${response.status}): ${url}`);\n  if (!response.body) throw new Error(`No response body: ${url}`);\n  const tempPath = `${destPath}.${process.pid}.${Date.now()}.tmp`;\n  try {\n    const fileStream = createWriteStream(tempPath);\n    await streamPipeline(Readable.fromWeb(response.body as never), fileStream);\n    await rename(tempPath, destPath);\n  } catch (err) {\n    try { await unlink(tempPath); } catch { /* temp may not exist */ }\n    throw err;\n  }\n}\n\nexport class Embedder {\n  private model: string;\n  private session: OrtSession | null = null;\n  private tokenizer: Awaited<ReturnType<typeof AutoTokenizer.from_pretrained>> | null = null;\n  private ort: OrtModule | null = null;\n  private dimensions = 0;\n  private runtimeLabel: RuntimeLabel = \"wasm\";\n  private initialized = false;\n  private numWorkers: number;\n  private batchSize: number;\n  private quantized: boolean;\n  private modelPath = \"\";\n\n  constructor(\n    model: string = DEFAULT_MODEL,\n    numWorkers: number = DEFAULT_WORKERS,\n    batchSize: number = DEFAULT_BATCH_SIZE,\n    quantized: boolean = DEFAULT_QUANTIZED\n  ) {\n    this.model = model;\n    this.numWorkers = numWorkers;\n    this.batchSize = batchSize;\n    this.quantized = quantized;\n  }\n\n  async init(): Promise<void> {\n    if (this.initialized) return;\n\n    const modelDir = join(CACHE_DIR, this.model.replace(/\\//g, \"--\"));\n    await mkdir(modelDir, { recursive: true });\n\n    // Resolve ONNX runtime (native C++ or WASM fallback)\n    const { ort, label } = await resolveOnnxRuntime();\n    this.ort = ort;\n    this.runtimeLabel = label;\n\n    // Download ONNX model if not cached. Tries quantized first if requested,\n    // falls back to fp32 if the quantized file is not available for this model.\n    let useQuantized = this.quantized;\n    const modelFileName = useQuantized ? \"model_quantized.onnx\" : \"model.onnx\";\n    this.modelPath = join(modelDir, modelFileName);\n    const modelPath = this.modelPath;\n    if (!existsSync(modelPath)) {\n      const pathMap = useQuantized ? ONNX_QUANTIZED_MODEL_PATHS : ONNX_MODEL_PATHS;\n      const onnxSubpath = pathMap[this.model] ?? (useQuantized ? \"onnx/model_quantized.onnx\" : \"onnx/model.onnx\");\n      const url = `https://huggingface.co/${this.model}/resolve/main/${onnxSubpath}`;\n      process.stderr.write(`Downloading ONNX model: ${this.model} (${useQuantized ? \"quantized\" : \"fp32\"})...\\n`);\n      try {\n        await downloadFile(url, modelPath);\n      } catch (err: any) {\n        if (useQuantized && err?.message?.includes(\"404\")) {\n          // Quantized not available for this model — fall back to fp32\n          process.stderr.write(`Quantized model not available, falling back to fp32\\n`);\n          useQuantized = false;\n          this.modelPath = join(modelDir, \"model.onnx\");\n          const fp32Subpath = ONNX_MODEL_PATHS[this.model] ?? \"onnx/model.onnx\";\n          const fp32Url = `https://huggingface.co/${this.model}/resolve/main/${fp32Subpath}`;\n          process.stderr.write(`Downloading ONNX model: ${this.model} (fp32)...\\n`);\n          await downloadFile(fp32Url, this.modelPath);\n        } else {\n          throw err;\n        }\n      }\n      process.stderr.write(`Model downloaded to ${modelDir}\\n`);\n    }\n    // Load tokenizer (uses HF transformers tokenizer infrastructure)\n    this.tokenizer = await AutoTokenizer.from_pretrained(this.model, {\n      cache_dir: CACHE_DIR,\n    });\n\n    // Create ONNX inference session (use this.modelPath — may have been updated by fallback)\n    this.session = await ort.InferenceSession.create(this.modelPath, {\n      executionProviders: [label === \"native\" ? \"cpu\" : \"wasm\"],\n    });\n\n    // Determine dimensions from a test embedding\n    const test = await this.embed(\"test\");\n    this.dimensions = test.length;\n    this.initialized = true;\n\n    const modelShort = this.model.split(\"/\").pop() ?? this.model;\n    process.stderr.write(`Embedder ready (${modelShort}, ${label}, ${this.dimensions}d, batch=${this.batchSize})\\n`);\n  }\n\n  async embed(text: string): Promise<Float32Array> {\n    if (!this.session || !this.tokenizer || !this.ort)\n      throw new Error(\"Embedder not initialized. Call init() first.\");\n\n    // Tokenize\n    const encoded = await this.tokenizer(text, {\n      padding: true,\n      truncation: true,\n      max_length: 512,\n      return_tensor: false,\n    });\n\n    const inputIdsRaw: number[] = Array.from(encoded.input_ids);\n    const attentionMaskRaw: number[] = Array.from(encoded.attention_mask);\n    const seqLen = inputIdsRaw.length;\n\n    // Build ONNX input tensors (most models expect int64)\n    const inputIds = new this.ort.Tensor(\n      \"int64\",\n      BigInt64Array.from(inputIdsRaw.map(BigInt)),\n      [1, seqLen]\n    );\n    const attentionMask = new this.ort.Tensor(\n      \"int64\",\n      BigInt64Array.from(attentionMaskRaw.map(BigInt)),\n      [1, seqLen]\n    );\n\n    const feeds: Record<string, unknown> = { input_ids: inputIds, attention_mask: attentionMask };\n\n    // Some models need token_type_ids\n    if (this.session.inputNames.includes(\"token_type_ids\")) {\n      feeds.token_type_ids = new this.ort.Tensor(\n        \"int64\",\n        new BigInt64Array(seqLen),\n        [1, seqLen]\n      );\n    }\n\n    // Run inference\n    const output = await this.session.run(feeds);\n    const outputTensor = output[this.session.outputNames[0]];\n    const hiddenSize = outputTensor.dims[outputTensor.dims.length - 1];\n\n    // Mean pooling with attention mask + L2 normalization\n    return this.meanPoolAndNormalize(outputTensor.data, attentionMaskRaw, seqLen, hiddenSize);\n  }\n\n  private meanPoolAndNormalize(\n    embeddings: Float32Array,\n    attentionMask: number[],\n    seqLen: number,\n    hiddenSize: number\n  ): Float32Array {\n    const result = new Float32Array(hiddenSize);\n    let maskSum = 0;\n\n    for (let t = 0; t < seqLen; t++) {\n      const mask = attentionMask[t];\n      maskSum += mask;\n      const offset = t * hiddenSize;\n      for (let d = 0; d < hiddenSize; d++) {\n        result[d] += embeddings[offset + d] * mask;\n      }\n    }\n\n    if (maskSum > 0) {\n      for (let d = 0; d < hiddenSize; d++) result[d] /= maskSum;\n    }\n\n    // L2 normalize\n    let norm = 0;\n    for (let d = 0; d < hiddenSize; d++) norm += result[d] * result[d];\n    norm = Math.sqrt(norm);\n    if (norm > 0) {\n      for (let d = 0; d < hiddenSize; d++) result[d] /= norm;\n    }\n\n    return result;\n  }\n\n  // Mean pool + L2 normalize the output of a batched ONNX forward pass.\n  // outputData: flat Float32Array of shape [batchSize, seqLen, hiddenSize]\n  // maskData:   flat number[] of shape [batchSize, seqLen]\n  private meanPoolAndNormalizeMany(\n    outputData: Float32Array,\n    maskData: number[],\n    batchSize: number,\n    seqLen: number,\n    hiddenSize: number\n  ): Float32Array[] {\n    const results: Float32Array[] = [];\n    for (let b = 0; b < batchSize; b++) {\n      const result = new Float32Array(hiddenSize);\n      let maskSum = 0;\n      for (let t = 0; t < seqLen; t++) {\n        const mask = maskData[b * seqLen + t];\n        maskSum += mask;\n        const offset = (b * seqLen + t) * hiddenSize;\n        for (let d = 0; d < hiddenSize; d++) {\n          result[d] += outputData[offset + d] * mask;\n        }\n      }\n      if (maskSum > 0) {\n        for (let d = 0; d < hiddenSize; d++) result[d] /= maskSum;\n      }\n      // L2 normalize\n      let norm = 0;\n      for (let d = 0; d < hiddenSize; d++) norm += result[d] * result[d];\n      norm = Math.sqrt(norm);\n      if (norm > 0) {\n        for (let d = 0; d < hiddenSize; d++) result[d] /= norm;\n      }\n      results.push(result);\n    }\n    return results;\n  }\n\n  // Embed a sub-batch of texts (length <= batchSize) in a single ONNX forward pass.\n  //\n  // Tokenizes each text individually (well-defined output format for all HF versions),\n  // then manually pads to the longest sequence in the batch and builds a [n, seqLen]\n  // tensor for one ONNX call. The speedup is from batching the ONNX inference —\n  // individual tokenization is negligible (<1ms each).\n  private async embedSubBatch(texts: string[]): Promise<Float32Array[]> {\n    if (!this.session || !this.tokenizer || !this.ort)\n      throw new Error(\"Embedder not initialized. Call init() first.\");\n\n    const n = texts.length;\n\n    // Tokenize each text individually — avoids ambiguous batch tokenizer output format\n    const encodings = await Promise.all(\n      texts.map((text) =>\n        this.tokenizer!(text, {\n          padding: false,\n          truncation: true,\n          max_length: 512,\n          return_tensor: false,\n        })\n      )\n    );\n\n    // Extract flat token arrays and find max sequence length for this batch\n    const tokenized = encodings.map((enc) => ({\n      ids: Array.from(enc.input_ids) as number[],\n      mask: Array.from(enc.attention_mask) as number[],\n    }));\n    const seqLen = Math.max(...tokenized.map((t) => t.ids.length));\n\n    // Build flat padded tensors [n * seqLen] — pad with 0 (PAD token, zero attention)\n    const flatIds = new BigInt64Array(n * seqLen);\n    const flatMask = new BigInt64Array(n * seqLen);\n    const flatMaskNums = new Array<number>(n * seqLen).fill(0);\n\n    for (let i = 0; i < n; i++) {\n      const { ids, mask } = tokenized[i];\n      for (let j = 0; j < ids.length; j++) {\n        flatIds[i * seqLen + j] = BigInt(ids[j]);\n        flatMask[i * seqLen + j] = BigInt(mask[j]);\n        flatMaskNums[i * seqLen + j] = mask[j];\n      }\n      // Positions beyond ids.length remain 0 (padding)\n    }\n\n    // Build batched ONNX tensors [n, seqLen]\n    const inputIds = new this.ort.Tensor(\"int64\", flatIds, [n, seqLen]);\n    const attentionMask = new this.ort.Tensor(\"int64\", flatMask, [n, seqLen]);\n\n    const feeds: Record<string, unknown> = { input_ids: inputIds, attention_mask: attentionMask };\n\n    if (this.session.inputNames.includes(\"token_type_ids\")) {\n      feeds.token_type_ids = new this.ort.Tensor(\n        \"int64\",\n        new BigInt64Array(n * seqLen),\n        [n, seqLen]\n      );\n    }\n\n    // Single forward pass → output shape [n, seqLen, hiddenSize]\n    const output = await this.session.run(feeds);\n    const outputTensor = output[this.session.outputNames[0]];\n    const hiddenSize = outputTensor.dims[outputTensor.dims.length - 1];\n\n    return this.meanPoolAndNormalizeMany(\n      outputTensor.data,\n      flatMaskNums,\n      n,\n      seqLen,\n      hiddenSize\n    );\n  }\n\n  async embedBatch(\n    texts: string[],\n    onProgress?: (embedded: number, total: number, subBatch?: Float32Array[]) => Promise<void> | void\n  ): Promise<Float32Array[]> {\n    // Workers path (disabled by default; kept for --workers N users)\n    if (this.numWorkers > 1 && texts.length >= this.numWorkers * 2) {\n      return this.embedBatchParallel(texts, onProgress);\n    }\n\n    // True batched inference: slice into sub-batches and run one ONNX call each\n    const results: Float32Array[] = [];\n    for (let i = 0; i < texts.length; i += this.batchSize) {\n      const subBatch = texts.slice(i, i + this.batchSize);\n      const embeddings = await this.embedSubBatch(subBatch);\n      results.push(...embeddings);\n      await onProgress?.(Math.min(i + subBatch.length, texts.length), texts.length, embeddings);\n    }\n    return results;\n  }\n\n  private async embedBatchParallel(\n    texts: string[],\n    onProgress?: (embedded: number, total: number, subBatch?: Float32Array[]) => Promise<void> | void\n  ): Promise<Float32Array[]> {\n    // Resolve worker script path. With tsup splitting, embedder code may land\n    // in a top-level chunk (dist/chunk-*.js) rather than dist/core/index.js,\n    // so we try two locations: adjacent to this chunk, and dist/core/ relative\n    // to the package root (two levels up from a dist/ chunk file).\n    const thisDir = dirname(fileURLToPath(import.meta.url));\n    let workerPath = join(thisDir, \"embed-worker.js\");\n    if (!existsSync(workerPath)) {\n      // Chunk file at dist/ level — worker is in dist/core/\n      workerPath = join(thisDir, \"core\", \"embed-worker.js\");\n    }\n\n    if (!existsSync(workerPath)) {\n      // Fallback to serial batched if worker script not found\n      process.stderr.write(\"Worker script not found, falling back to batched embedding\\n\");\n      return this.embedBatch(texts, onProgress);\n    }\n\n    // Split texts into chunks for each worker\n    const chunkSize = Math.ceil(texts.length / this.numWorkers);\n    const chunks: { texts: string[]; startIndex: number }[] = [];\n    for (let i = 0; i < texts.length; i += chunkSize) {\n      chunks.push({ texts: texts.slice(i, i + chunkSize), startIndex: i });\n    }\n\n    const allResults = new Array<Float32Array>(texts.length);\n    let totalDone = 0;\n\n    const workerPromises = chunks.map((chunk) => {\n      return new Promise<void>((resolve, reject) => {\n        const worker = new Worker(workerPath, {\n          workerData: {\n            modelPath: this.modelPath,\n            modelName: this.model,\n            cacheDir: CACHE_DIR,\n            runtimeLabel: this.runtimeLabel,\n            batchSize: this.batchSize,\n          },\n        });\n\n        worker.on(\"message\", (msg: any) => {\n          if (msg.type === \"ready\") {\n            worker.postMessage({ type: \"embed\", texts: chunk.texts, startIndex: chunk.startIndex });\n          } else if (msg.type === \"progress\") {\n            totalDone++;\n            onProgress?.(totalDone, texts.length);\n          } else if (msg.type === \"result\") {\n            for (let i = 0; i < msg.embeddings.length; i++) {\n              allResults[chunk.startIndex + i] = new Float32Array(msg.embeddings[i]);\n            }\n            worker.terminate();\n            resolve();\n          } else if (msg.type === \"error\") {\n            worker.terminate();\n            reject(new Error(msg.error));\n          }\n        });\n\n        worker.on(\"error\", (err) => {\n          worker.terminate();\n          reject(err);\n        });\n      });\n    });\n\n    await Promise.all(workerPromises);\n    return allResults;\n  }\n\n  getDimensions(): number {\n    return this.dimensions;\n  }\n\n  getModel(): string {\n    return this.model;\n  }\n\n  getRuntime(): RuntimeLabel {\n    return this.runtimeLabel;\n  }\n\n  async saveEmbeddings(\n    embeddings: Map<string, Float32Array>,\n    indexPath: string\n  ): Promise<void> {\n    const entries: Array<{ key: string; data: number[] }> = [];\n    for (const [key, vec] of embeddings) {\n      entries.push({ key, data: Array.from(vec) });\n    }\n    await writeFile(join(indexPath, \"embeddings.json\"), JSON.stringify(entries));\n  }\n\n  async loadEmbeddings(\n    indexPath: string\n  ): Promise<Map<string, Float32Array>> {\n    const filePath = join(indexPath, \"embeddings.json\");\n    if (!existsSync(filePath)) return new Map();\n\n    const raw = await readFile(filePath, \"utf-8\");\n    const entries: Array<{ key: string; data: number[] }> = JSON.parse(raw);\n    const map = new Map<string, Float32Array>();\n    for (const entry of entries) {\n      map.set(entry.key, new Float32Array(entry.data));\n    }\n    return map;\n  }\n}\n","import Graph from \"graphology\";\nimport { bfsFromNode } from \"graphology-traversal\";\nimport { bidirectional } from \"graphology-shortest-path\";\nimport type { IndexedDocument, GraphNode, GraphEdge, GraphStats } from \"./types.js\";\nimport { readFile, writeFile } from \"node:fs/promises\";\nimport { join } from \"node:path\";\nimport { existsSync } from \"node:fs\";\n\n/**\n * Tags shared by more than this many docs are too generic to be meaningful\n * graph edges (e.g. \"api\", \"testing\"). Only specific/rare tags create edges.\n * Docs without frontmatter are unaffected — their inline #tags still participate\n * but only if those tags are specific enough to pass this threshold.\n */\nconst TAG_EDGE_MAX_DOCS = 15;\n\nexport class GraphBuilder {\n  private graph: Graph;\n\n  constructor() {\n    this.graph = new Graph({ type: \"directed\", multi: false });\n  }\n\n  buildFromDocuments(documents: IndexedDocument[]): void {\n    this.graph.clear();\n\n    // Add nodes — store optional metadata fields when present\n    for (const doc of documents) {\n      this.graph.addNode(doc.path, {\n        title: doc.title,\n        tags: doc.tags,\n        ...(doc.status !== undefined && { status: doc.status }),\n        ...(doc.loadPriority !== undefined && { loadPriority: doc.loadPriority }),\n      });\n    }\n\n    const pathLookup = new Map<string, string>();\n    for (const doc of documents) {\n      const nameNoExt = doc.path.replace(/\\.md$/, \"\");\n      const basename = nameNoExt.split(\"/\").pop()!;\n      pathLookup.set(basename.toLowerCase(), doc.path);\n      pathLookup.set(nameNoExt.toLowerCase(), doc.path);\n    }\n\n    // Add wikilink edges (weight 1.0)\n    for (const doc of documents) {\n      for (const link of doc.wikilinks) {\n        const target = pathLookup.get(link.toLowerCase());\n        if (target && target !== doc.path && !this.graph.hasEdge(doc.path, target)) {\n          this.graph.addEdge(doc.path, target, { type: \"wikilink\", weight: 1.0 });\n        }\n      }\n    }\n\n    // Add related_docs edges (weight 1.0) — only present when frontmatter has related_docs\n    // Falls back gracefully: docs without this field are simply skipped\n    for (const doc of documents) {\n      if (!doc.relatedDocs?.length) continue;\n      for (const ref of doc.relatedDocs) {\n        // Try full normalized path first, then basename — same strategy as wikilinks\n        const normalized = ref.replace(/\\.md$/, \"\").toLowerCase();\n        const base = normalized.split(\"/\").pop()!;\n        const target = pathLookup.get(normalized) ?? pathLookup.get(base);\n        if (target && target !== doc.path && !this.graph.hasEdge(doc.path, target)) {\n          this.graph.addEdge(doc.path, target, { type: \"related\", weight: 1.0 });\n        }\n      }\n    }\n\n    // Add tag-based edges (weight 0.5) — capped at TAG_EDGE_MAX_DOCS per tag.\n    // Generic tags shared by many docs (e.g. \"api\", \"testing\") create noise;\n    // only specific tags (e.g. \"qdrant\", \"celery\") create meaningful connections.\n    const tagToNotes = new Map<string, string[]>();\n    for (const doc of documents) {\n      for (const tag of doc.tags) {\n        const notes = tagToNotes.get(tag) ?? [];\n        notes.push(doc.path);\n        tagToNotes.set(tag, notes);\n      }\n    }\n\n    for (const [, notes] of tagToNotes) {\n      if (notes.length >= TAG_EDGE_MAX_DOCS) continue; // too generic — skip\n      for (let i = 0; i < notes.length; i++) {\n        for (let j = i + 1; j < notes.length; j++) {\n          if (!this.graph.hasEdge(notes[i], notes[j])) {\n            this.graph.addEdge(notes[i], notes[j], { type: \"tag\", weight: 0.5 });\n          }\n          if (!this.graph.hasEdge(notes[j], notes[i])) {\n            this.graph.addEdge(notes[j], notes[i], { type: \"tag\", weight: 0.5 });\n          }\n        }\n      }\n    }\n  }\n\n  backlinks(notePath: string): GraphNode[] {\n    if (!this.graph.hasNode(notePath)) return [];\n    return this.graph.inNeighbors(notePath).map((n) => this.nodeToGraphNode(n));\n  }\n\n  forwardlinks(notePath: string): GraphNode[] {\n    if (!this.graph.hasNode(notePath)) return [];\n    return this.graph.outNeighbors(notePath).map((n) => this.nodeToGraphNode(n));\n  }\n\n  findPath(from: string, to: string): string[] | null {\n    if (!this.graph.hasNode(from) || !this.graph.hasNode(to)) return null;\n    const path = bidirectional(this.graph, from, to);\n    return path;\n  }\n\n  searchGraph(concept: string, maxDepth = 2, limit = 20): GraphNode[] {\n    const lc = concept.toLowerCase();\n\n    const directMatches = new Set<string>(\n      this.graph.nodes().filter((n) => {\n        const attrs = this.graph.getNodeAttributes(n);\n        const title = typeof attrs.title === \"string\" ? attrs.title : String(attrs.title ?? \"\");\n        return (\n          n.toLowerCase().includes(lc) ||\n          title.toLowerCase().includes(lc) ||\n          (attrs.tags as string[] | undefined)?.some((t) => t.toLowerCase().includes(lc))\n        );\n      })\n    );\n\n    const visited = new Set<string>();\n    for (const start of directMatches) {\n      bfsFromNode(this.graph, start, (_node, _attr, depth) => {\n        if (visited.size >= limit * 3) return true; // gather 3x limit before ranking\n        visited.add(_node);\n        return depth >= maxDepth;\n      });\n    }\n\n    // Sort: direct matches first, then by loadPriority desc, then by degree desc\n    // deprecated/archived nodes sort to the end\n    const ranked = [...visited].sort((a, b) => {\n      const aAttrs = this.graph.getNodeAttributes(a);\n      const bAttrs = this.graph.getNodeAttributes(b);\n      const aDeprecated = aAttrs.status === \"deprecated\" || aAttrs.status === \"archived\" ? 1 : 0;\n      const bDeprecated = bAttrs.status === \"deprecated\" || bAttrs.status === \"archived\" ? 1 : 0;\n      if (aDeprecated !== bDeprecated) return aDeprecated - bDeprecated;\n      const aDirect = directMatches.has(a) ? 0 : 1;\n      const bDirect = directMatches.has(b) ? 0 : 1;\n      if (aDirect !== bDirect) return aDirect - bDirect;\n      const aPriority = (aAttrs.loadPriority as number | undefined) ?? 5;\n      const bPriority = (bAttrs.loadPriority as number | undefined) ?? 5;\n      if (aPriority !== bPriority) return bPriority - aPriority;\n      return this.graph.degree(b) - this.graph.degree(a);\n    });\n\n    return ranked.slice(0, limit).map((n) => this.nodeToGraphNode(n));\n  }\n\n  statistics(): GraphStats {\n    const nodes = this.graph.order;\n    const edges = this.graph.size;\n    const orphans = this.graph.nodes().filter(\n      (n) => this.graph.degree(n) === 0\n    );\n\n    const connections = this.graph.nodes().map((n) => ({\n      path: n,\n      connections: this.graph.degree(n),\n    }));\n    connections.sort((a, b) => b.connections - a.connections);\n\n    const maxPossibleEdges = nodes * (nodes - 1);\n    const density = maxPossibleEdges > 0 ? edges / maxPossibleEdges : 0;\n\n    return {\n      totalNodes: nodes,\n      totalEdges: edges,\n      orphanCount: orphans.length,\n      mostConnected: connections.slice(0, 10),\n      density,\n    };\n  }\n\n  private nodeToGraphNode(nodePath: string): GraphNode {\n    const attrs = this.graph.getNodeAttributes(nodePath);\n    return {\n      path: nodePath,\n      title: typeof attrs.title === \"string\" ? attrs.title : nodePath,\n      tags: (attrs.tags as string[]) || [],\n      linkCount: this.graph.outDegree(nodePath),\n      backlinkCount: this.graph.inDegree(nodePath),\n      ...(attrs.loadPriority !== undefined && { loadPriority: attrs.loadPriority as number }),\n      ...(attrs.status !== undefined && { status: attrs.status as string }),\n    };\n  }\n\n  async save(indexPath: string): Promise<void> {\n    const data = this.graph.export();\n    await writeFile(join(indexPath, \"graph.json\"), JSON.stringify(data));\n  }\n\n  async load(indexPath: string): Promise<boolean> {\n    const filePath = join(indexPath, \"graph.json\");\n    if (!existsSync(filePath)) return false;\n\n    const raw = await readFile(filePath, \"utf-8\");\n    const data = JSON.parse(raw);\n    this.graph.import(data);\n    return true;\n  }\n\n  getGraph(): Graph {\n    return this.graph;\n  }\n}\n","import hnswlib from \"hnswlib-node\";\nconst { HierarchicalNSW } = hnswlib;\nimport { readFile, writeFile } from \"node:fs/promises\";\nimport { join } from \"node:path\";\nimport { existsSync } from \"node:fs\";\nimport type { SearchResult } from \"./types.js\";\n\ninterface ChunkMeta {\n  docPath: string;\n  chunkIndex: number;\n  text: string;\n}\n\nexport class VectorIndex {\n  private index: InstanceType<typeof HierarchicalNSW> | null = null;\n  private dimensions: number;\n  private chunkMeta: ChunkMeta[] = [];\n\n  constructor(dimensions: number) {\n    this.dimensions = dimensions;\n  }\n\n  build(\n    embeddings: Float32Array[],\n    meta: ChunkMeta[]\n  ): void {\n    if (embeddings.length === 0) {\n      this.index = null;\n      this.chunkMeta = [];\n      return;\n    }\n\n    this.index = new HierarchicalNSW(\"cosine\", this.dimensions);\n    this.index.initIndex(embeddings.length);\n\n    for (let i = 0; i < embeddings.length; i++) {\n      this.index.addPoint(Array.from(embeddings[i]), i);\n    }\n\n    this.chunkMeta = meta;\n  }\n\n  search(queryEmbedding: Float32Array, k: number = 10): SearchResult[] {\n    if (!this.index || this.chunkMeta.length === 0) return [];\n\n    const numResults = Math.min(k, this.chunkMeta.length);\n    const result = this.index.searchKnn(Array.from(queryEmbedding), numResults);\n\n    const seen = new Set<string>();\n    const results: SearchResult[] = [];\n\n    for (let i = 0; i < result.neighbors.length; i++) {\n      const idx = result.neighbors[i];\n      const meta = this.chunkMeta[idx];\n      if (!meta || seen.has(meta.docPath)) continue;\n      seen.add(meta.docPath);\n\n      results.push({\n        path: meta.docPath,\n        title: meta.docPath,\n        score: 1 - result.distances[i],\n        snippet: meta.text.slice(0, 200),\n        matchedChunk: meta.text,\n      });\n    }\n\n    return results;\n  }\n\n  async save(indexPath: string): Promise<void> {\n    if (!this.index) return;\n\n    this.index.writeIndexSync(join(indexPath, \"hnsw.bin\"));\n    await writeFile(\n      join(indexPath, \"hnsw-meta.json\"),\n      JSON.stringify(this.chunkMeta)\n    );\n  }\n\n  async load(indexPath: string): Promise<boolean> {\n    const hnswPath = join(indexPath, \"hnsw.bin\");\n    const metaPath = join(indexPath, \"hnsw-meta.json\");\n\n    if (!existsSync(hnswPath) || !existsSync(metaPath)) return false;\n\n    const raw = await readFile(metaPath, \"utf-8\");\n    this.chunkMeta = JSON.parse(raw);\n\n    this.index = new HierarchicalNSW(\"cosine\", this.dimensions);\n    this.index.initIndex(this.chunkMeta.length);\n    this.index.readIndexSync(hnswPath);\n\n    return true;\n  }\n\n  getChunkMeta(): ChunkMeta[] {\n    return this.chunkMeta;\n  }\n}\n","import type { IndexedDocument, SearchResult, SearchTextOptions } from \"./types.js\";\nimport { minimatch } from \"minimatch\";\n\nexport class TextSearch {\n  private documents: IndexedDocument[] = [];\n\n  setDocuments(documents: IndexedDocument[]): void {\n    this.documents = documents;\n  }\n\n  search(options: SearchTextOptions): SearchResult[] {\n    const { pattern, regex, caseSensitive, pathGlob, tagFilter, limit = 20 } = options;\n\n    let matcher: (text: string) => { matched: boolean; index: number };\n\n    if (regex) {\n      const flags = caseSensitive ? \"g\" : \"gi\";\n      const re = new RegExp(pattern, flags);\n      matcher = (text) => {\n        re.lastIndex = 0;\n        const m = re.exec(text);\n        return { matched: !!m, index: m?.index ?? -1 };\n      };\n    } else {\n      const needle = caseSensitive ? pattern : pattern.toLowerCase();\n      matcher = (text) => {\n        const haystack = caseSensitive ? text : text.toLowerCase();\n        const idx = haystack.indexOf(needle);\n        return { matched: idx >= 0, index: idx };\n      };\n    }\n\n    const results: SearchResult[] = [];\n\n    for (const doc of this.documents) {\n      // Path filter\n      if (pathGlob && !minimatch(doc.path, pathGlob)) continue;\n\n      // Tag filter\n      if (tagFilter?.length) {\n        const hasTag = tagFilter.some((t) => doc.tags.includes(t));\n        if (!hasTag) continue;\n      }\n\n      const { matched, index } = matcher(doc.content);\n      if (!matched) continue;\n\n      const snippetStart = Math.max(0, index - 80);\n      const snippetEnd = Math.min(doc.content.length, index + 120);\n      const snippet = doc.content.slice(snippetStart, snippetEnd).trim();\n\n      results.push({\n        path: doc.path,\n        title: doc.title,\n        score: 1.0,\n        snippet: (snippetStart > 0 ? \"...\" : \"\") + snippet + (snippetEnd < doc.content.length ? \"...\" : \"\"),\n      });\n\n      if (results.length >= limit) break;\n    }\n\n    return results;\n  }\n}\n"],"mappings":";AAAA,SAAS,qBAAqB;AAC9B,SAAS,UAAU,WAAW,OAAO,QAAQ,cAAc;AAC3D,SAAS,MAAM,eAAe;AAC9B,SAAS,YAAY,yBAAyB;AAC9C,SAAS,eAAe;AACxB,SAAS,gBAAgB;AACzB,SAAS,YAAY,sBAAsB;AAC3C,SAAS,cAAc;AACvB,SAAS,qBAAqB;AAI9B,IAAM,gBAAgB;AACtB,IAAM,YAAY,KAAK,QAAQ,GAAG,sBAAsB,QAAQ;AAKhE,IAAM,kBAAkB;AAGxB,IAAM,qBAAqB;AAI3B,IAAM,oBAAoB;AAG1B,IAAM,mBAA2C;AAAA,EAC/C,kCAAkC;AAAA,EAClC,0CAA0C;AAC5C;AAGA,IAAM,6BAAqD;AAAA,EACzD,kCAAkC;AAAA,EAClC,0CAA0C;AAC5C;AAiBA,eAAe,qBAAuE;AACpF,MAAI;AACF,UAAM,MAAM,MAAM,OAAO,kBAAkB;AAC3C,WAAO,EAAE,KAAkC,OAAO,SAAS;AAAA,EAC7D,QAAQ;AACN,UAAM,MAAM,MAAM,OAAO,iBAAiB;AAC1C,WAAO,EAAE,KAAkC,OAAO,OAAO;AAAA,EAC3D;AACF;AAUA,eAAe,aAAa,KAAa,UAAiC;AACxE,QAAM,WAAW,MAAM,MAAM,GAAG;AAChC,MAAI,CAAC,SAAS,GAAI,OAAM,IAAI,MAAM,oBAAoB,SAAS,MAAM,MAAM,GAAG,EAAE;AAChF,MAAI,CAAC,SAAS,KAAM,OAAM,IAAI,MAAM,qBAAqB,GAAG,EAAE;AAC9D,QAAM,WAAW,GAAG,QAAQ,IAAI,QAAQ,GAAG,IAAI,KAAK,IAAI,CAAC;AACzD,MAAI;AACF,UAAM,aAAa,kBAAkB,QAAQ;AAC7C,UAAM,eAAe,SAAS,QAAQ,SAAS,IAAa,GAAG,UAAU;AACzE,UAAM,OAAO,UAAU,QAAQ;AAAA,EACjC,SAAS,KAAK;AACZ,QAAI;AAAE,YAAM,OAAO,QAAQ;AAAA,IAAG,QAAQ;AAAA,IAA2B;AACjE,UAAM;AAAA,EACR;AACF;AAEO,IAAM,WAAN,MAAe;AAAA,EACZ;AAAA,EACA,UAA6B;AAAA,EAC7B,YAA8E;AAAA,EAC9E,MAAwB;AAAA,EACxB,aAAa;AAAA,EACb,eAA6B;AAAA,EAC7B,cAAc;AAAA,EACd;AAAA,EACA;AAAA,EACA;AAAA,EACA,YAAY;AAAA,EAEpB,YACE,QAAgB,eAChB,aAAqB,iBACrB,YAAoB,oBACpB,YAAqB,mBACrB;AACA,SAAK,QAAQ;AACb,SAAK,aAAa;AAClB,SAAK,YAAY;AACjB,SAAK,YAAY;AAAA,EACnB;AAAA,EAEA,MAAM,OAAsB;AAC1B,QAAI,KAAK,YAAa;AAEtB,UAAM,WAAW,KAAK,WAAW,KAAK,MAAM,QAAQ,OAAO,IAAI,CAAC;AAChE,UAAM,MAAM,UAAU,EAAE,WAAW,KAAK,CAAC;AAGzC,UAAM,EAAE,KAAK,MAAM,IAAI,MAAM,mBAAmB;AAChD,SAAK,MAAM;AACX,SAAK,eAAe;AAIpB,QAAI,eAAe,KAAK;AACxB,UAAM,gBAAgB,eAAe,yBAAyB;AAC9D,SAAK,YAAY,KAAK,UAAU,aAAa;AAC7C,UAAM,YAAY,KAAK;AACvB,QAAI,CAAC,WAAW,SAAS,GAAG;AAC1B,YAAM,UAAU,eAAe,6BAA6B;AAC5D,YAAM,cAAc,QAAQ,KAAK,KAAK,MAAM,eAAe,8BAA8B;AACzF,YAAM,MAAM,0BAA0B,KAAK,KAAK,iBAAiB,WAAW;AAC5E,cAAQ,OAAO,MAAM,2BAA2B,KAAK,KAAK,KAAK,eAAe,cAAc,MAAM;AAAA,CAAQ;AAC1G,UAAI;AACF,cAAM,aAAa,KAAK,SAAS;AAAA,MACnC,SAAS,KAAU;AACjB,YAAI,gBAAgB,KAAK,SAAS,SAAS,KAAK,GAAG;AAEjD,kBAAQ,OAAO,MAAM;AAAA,CAAuD;AAC5E,yBAAe;AACf,eAAK,YAAY,KAAK,UAAU,YAAY;AAC5C,gBAAM,cAAc,iBAAiB,KAAK,KAAK,KAAK;AACpD,gBAAM,UAAU,0BAA0B,KAAK,KAAK,iBAAiB,WAAW;AAChF,kBAAQ,OAAO,MAAM,2BAA2B,KAAK,KAAK;AAAA,CAAc;AACxE,gBAAM,aAAa,SAAS,KAAK,SAAS;AAAA,QAC5C,OAAO;AACL,gBAAM;AAAA,QACR;AAAA,MACF;AACA,cAAQ,OAAO,MAAM,uBAAuB,QAAQ;AAAA,CAAI;AAAA,IAC1D;AAEA,SAAK,YAAY,MAAM,cAAc,gBAAgB,KAAK,OAAO;AAAA,MAC/D,WAAW;AAAA,IACb,CAAC;AAGD,SAAK,UAAU,MAAM,IAAI,iBAAiB,OAAO,KAAK,WAAW;AAAA,MAC/D,oBAAoB,CAAC,UAAU,WAAW,QAAQ,MAAM;AAAA,IAC1D,CAAC;AAGD,UAAM,OAAO,MAAM,KAAK,MAAM,MAAM;AACpC,SAAK,aAAa,KAAK;AACvB,SAAK,cAAc;AAEnB,UAAM,aAAa,KAAK,MAAM,MAAM,GAAG,EAAE,IAAI,KAAK,KAAK;AACvD,YAAQ,OAAO,MAAM,mBAAmB,UAAU,KAAK,KAAK,KAAK,KAAK,UAAU,YAAY,KAAK,SAAS;AAAA,CAAK;AAAA,EACjH;AAAA,EAEA,MAAM,MAAM,MAAqC;AAC/C,QAAI,CAAC,KAAK,WAAW,CAAC,KAAK,aAAa,CAAC,KAAK;AAC5C,YAAM,IAAI,MAAM,8CAA8C;AAGhE,UAAM,UAAU,MAAM,KAAK,UAAU,MAAM;AAAA,MACzC,SAAS;AAAA,MACT,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,eAAe;AAAA,IACjB,CAAC;AAED,UAAM,cAAwB,MAAM,KAAK,QAAQ,SAAS;AAC1D,UAAM,mBAA6B,MAAM,KAAK,QAAQ,cAAc;AACpE,UAAM,SAAS,YAAY;AAG3B,UAAM,WAAW,IAAI,KAAK,IAAI;AAAA,MAC5B;AAAA,MACA,cAAc,KAAK,YAAY,IAAI,MAAM,CAAC;AAAA,MAC1C,CAAC,GAAG,MAAM;AAAA,IACZ;AACA,UAAM,gBAAgB,IAAI,KAAK,IAAI;AAAA,MACjC;AAAA,MACA,cAAc,KAAK,iBAAiB,IAAI,MAAM,CAAC;AAAA,MAC/C,CAAC,GAAG,MAAM;AAAA,IACZ;AAEA,UAAM,QAAiC,EAAE,WAAW,UAAU,gBAAgB,cAAc;AAG5F,QAAI,KAAK,QAAQ,WAAW,SAAS,gBAAgB,GAAG;AACtD,YAAM,iBAAiB,IAAI,KAAK,IAAI;AAAA,QAClC;AAAA,QACA,IAAI,cAAc,MAAM;AAAA,QACxB,CAAC,GAAG,MAAM;AAAA,MACZ;AAAA,IACF;AAGA,UAAM,SAAS,MAAM,KAAK,QAAQ,IAAI,KAAK;AAC3C,UAAM,eAAe,OAAO,KAAK,QAAQ,YAAY,CAAC,CAAC;AACvD,UAAM,aAAa,aAAa,KAAK,aAAa,KAAK,SAAS,CAAC;AAGjE,WAAO,KAAK,qBAAqB,aAAa,MAAM,kBAAkB,QAAQ,UAAU;AAAA,EAC1F;AAAA,EAEQ,qBACN,YACA,eACA,QACA,YACc;AACd,UAAM,SAAS,IAAI,aAAa,UAAU;AAC1C,QAAI,UAAU;AAEd,aAAS,IAAI,GAAG,IAAI,QAAQ,KAAK;AAC/B,YAAM,OAAO,cAAc,CAAC;AAC5B,iBAAW;AACX,YAAM,SAAS,IAAI;AACnB,eAAS,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,eAAO,CAAC,KAAK,WAAW,SAAS,CAAC,IAAI;AAAA,MACxC;AAAA,IACF;AAEA,QAAI,UAAU,GAAG;AACf,eAAS,IAAI,GAAG,IAAI,YAAY,IAAK,QAAO,CAAC,KAAK;AAAA,IACpD;AAGA,QAAI,OAAO;AACX,aAAS,IAAI,GAAG,IAAI,YAAY,IAAK,SAAQ,OAAO,CAAC,IAAI,OAAO,CAAC;AACjE,WAAO,KAAK,KAAK,IAAI;AACrB,QAAI,OAAO,GAAG;AACZ,eAAS,IAAI,GAAG,IAAI,YAAY,IAAK,QAAO,CAAC,KAAK;AAAA,IACpD;AAEA,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA,EAKQ,yBACN,YACA,UACA,WACA,QACA,YACgB;AAChB,UAAM,UAA0B,CAAC;AACjC,aAAS,IAAI,GAAG,IAAI,WAAW,KAAK;AAClC,YAAM,SAAS,IAAI,aAAa,UAAU;AAC1C,UAAI,UAAU;AACd,eAAS,IAAI,GAAG,IAAI,QAAQ,KAAK;AAC/B,cAAM,OAAO,SAAS,IAAI,SAAS,CAAC;AACpC,mBAAW;AACX,cAAM,UAAU,IAAI,SAAS,KAAK;AAClC,iBAAS,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,iBAAO,CAAC,KAAK,WAAW,SAAS,CAAC,IAAI;AAAA,QACxC;AAAA,MACF;AACA,UAAI,UAAU,GAAG;AACf,iBAAS,IAAI,GAAG,IAAI,YAAY,IAAK,QAAO,CAAC,KAAK;AAAA,MACpD;AAEA,UAAI,OAAO;AACX,eAAS,IAAI,GAAG,IAAI,YAAY,IAAK,SAAQ,OAAO,CAAC,IAAI,OAAO,CAAC;AACjE,aAAO,KAAK,KAAK,IAAI;AACrB,UAAI,OAAO,GAAG;AACZ,iBAAS,IAAI,GAAG,IAAI,YAAY,IAAK,QAAO,CAAC,KAAK;AAAA,MACpD;AACA,cAAQ,KAAK,MAAM;AAAA,IACrB;AACA,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAc,cAAc,OAA0C;AACpE,QAAI,CAAC,KAAK,WAAW,CAAC,KAAK,aAAa,CAAC,KAAK;AAC5C,YAAM,IAAI,MAAM,8CAA8C;AAEhE,UAAM,IAAI,MAAM;AAGhB,UAAM,YAAY,MAAM,QAAQ;AAAA,MAC9B,MAAM;AAAA,QAAI,CAAC,SACT,KAAK,UAAW,MAAM;AAAA,UACpB,SAAS;AAAA,UACT,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,eAAe;AAAA,QACjB,CAAC;AAAA,MACH;AAAA,IACF;AAGA,UAAM,YAAY,UAAU,IAAI,CAAC,SAAS;AAAA,MACxC,KAAK,MAAM,KAAK,IAAI,SAAS;AAAA,MAC7B,MAAM,MAAM,KAAK,IAAI,cAAc;AAAA,IACrC,EAAE;AACF,UAAM,SAAS,KAAK,IAAI,GAAG,UAAU,IAAI,CAAC,MAAM,EAAE,IAAI,MAAM,CAAC;AAG7D,UAAM,UAAU,IAAI,cAAc,IAAI,MAAM;AAC5C,UAAM,WAAW,IAAI,cAAc,IAAI,MAAM;AAC7C,UAAM,eAAe,IAAI,MAAc,IAAI,MAAM,EAAE,KAAK,CAAC;AAEzD,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAM,EAAE,KAAK,KAAK,IAAI,UAAU,CAAC;AACjC,eAAS,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;AACnC,gBAAQ,IAAI,SAAS,CAAC,IAAI,OAAO,IAAI,CAAC,CAAC;AACvC,iBAAS,IAAI,SAAS,CAAC,IAAI,OAAO,KAAK,CAAC,CAAC;AACzC,qBAAa,IAAI,SAAS,CAAC,IAAI,KAAK,CAAC;AAAA,MACvC;AAAA,IAEF;AAGA,UAAM,WAAW,IAAI,KAAK,IAAI,OAAO,SAAS,SAAS,CAAC,GAAG,MAAM,CAAC;AAClE,UAAM,gBAAgB,IAAI,KAAK,IAAI,OAAO,SAAS,UAAU,CAAC,GAAG,MAAM,CAAC;AAExE,UAAM,QAAiC,EAAE,WAAW,UAAU,gBAAgB,cAAc;AAE5F,QAAI,KAAK,QAAQ,WAAW,SAAS,gBAAgB,GAAG;AACtD,YAAM,iBAAiB,IAAI,KAAK,IAAI;AAAA,QAClC;AAAA,QACA,IAAI,cAAc,IAAI,MAAM;AAAA,QAC5B,CAAC,GAAG,MAAM;AAAA,MACZ;AAAA,IACF;AAGA,UAAM,SAAS,MAAM,KAAK,QAAQ,IAAI,KAAK;AAC3C,UAAM,eAAe,OAAO,KAAK,QAAQ,YAAY,CAAC,CAAC;AACvD,UAAM,aAAa,aAAa,KAAK,aAAa,KAAK,SAAS,CAAC;AAEjE,WAAO,KAAK;AAAA,MACV,aAAa;AAAA,MACb;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA,EAEA,MAAM,WACJ,OACA,YACyB;AAEzB,QAAI,KAAK,aAAa,KAAK,MAAM,UAAU,KAAK,aAAa,GAAG;AAC9D,aAAO,KAAK,mBAAmB,OAAO,UAAU;AAAA,IAClD;AAGA,UAAM,UAA0B,CAAC;AACjC,aAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK,KAAK,WAAW;AACrD,YAAM,WAAW,MAAM,MAAM,GAAG,IAAI,KAAK,SAAS;AAClD,YAAM,aAAa,MAAM,KAAK,cAAc,QAAQ;AACpD,cAAQ,KAAK,GAAG,UAAU;AAC1B,YAAM,aAAa,KAAK,IAAI,IAAI,SAAS,QAAQ,MAAM,MAAM,GAAG,MAAM,QAAQ,UAAU;AAAA,IAC1F;AACA,WAAO;AAAA,EACT;AAAA,EAEA,MAAc,mBACZ,OACA,YACyB;AAKzB,UAAM,UAAU,QAAQ,cAAc,YAAY,GAAG,CAAC;AACtD,QAAI,aAAa,KAAK,SAAS,iBAAiB;AAChD,QAAI,CAAC,WAAW,UAAU,GAAG;AAE3B,mBAAa,KAAK,SAAS,QAAQ,iBAAiB;AAAA,IACtD;AAEA,QAAI,CAAC,WAAW,UAAU,GAAG;AAE3B,cAAQ,OAAO,MAAM,8DAA8D;AACnF,aAAO,KAAK,WAAW,OAAO,UAAU;AAAA,IAC1C;AAGA,UAAM,YAAY,KAAK,KAAK,MAAM,SAAS,KAAK,UAAU;AAC1D,UAAM,SAAoD,CAAC;AAC3D,aAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK,WAAW;AAChD,aAAO,KAAK,EAAE,OAAO,MAAM,MAAM,GAAG,IAAI,SAAS,GAAG,YAAY,EAAE,CAAC;AAAA,IACrE;AAEA,UAAM,aAAa,IAAI,MAAoB,MAAM,MAAM;AACvD,QAAI,YAAY;AAEhB,UAAM,iBAAiB,OAAO,IAAI,CAAC,UAAU;AAC3C,aAAO,IAAI,QAAc,CAAC,SAAS,WAAW;AAC5C,cAAM,SAAS,IAAI,OAAO,YAAY;AAAA,UACpC,YAAY;AAAA,YACV,WAAW,KAAK;AAAA,YAChB,WAAW,KAAK;AAAA,YAChB,UAAU;AAAA,YACV,cAAc,KAAK;AAAA,YACnB,WAAW,KAAK;AAAA,UAClB;AAAA,QACF,CAAC;AAED,eAAO,GAAG,WAAW,CAAC,QAAa;AACjC,cAAI,IAAI,SAAS,SAAS;AACxB,mBAAO,YAAY,EAAE,MAAM,SAAS,OAAO,MAAM,OAAO,YAAY,MAAM,WAAW,CAAC;AAAA,UACxF,WAAW,IAAI,SAAS,YAAY;AAClC;AACA,yBAAa,WAAW,MAAM,MAAM;AAAA,UACtC,WAAW,IAAI,SAAS,UAAU;AAChC,qBAAS,IAAI,GAAG,IAAI,IAAI,WAAW,QAAQ,KAAK;AAC9C,yBAAW,MAAM,aAAa,CAAC,IAAI,IAAI,aAAa,IAAI,WAAW,CAAC,CAAC;AAAA,YACvE;AACA,mBAAO,UAAU;AACjB,oBAAQ;AAAA,UACV,WAAW,IAAI,SAAS,SAAS;AAC/B,mBAAO,UAAU;AACjB,mBAAO,IAAI,MAAM,IAAI,KAAK,CAAC;AAAA,UAC7B;AAAA,QACF,CAAC;AAED,eAAO,GAAG,SAAS,CAAC,QAAQ;AAC1B,iBAAO,UAAU;AACjB,iBAAO,GAAG;AAAA,QACZ,CAAC;AAAA,MACH,CAAC;AAAA,IACH,CAAC;AAED,UAAM,QAAQ,IAAI,cAAc;AAChC,WAAO;AAAA,EACT;AAAA,EAEA,gBAAwB;AACtB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,WAAmB;AACjB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,aAA2B;AACzB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,eACJ,YACA,WACe;AACf,UAAM,UAAkD,CAAC;AACzD,eAAW,CAAC,KAAK,GAAG,KAAK,YAAY;AACnC,cAAQ,KAAK,EAAE,KAAK,MAAM,MAAM,KAAK,GAAG,EAAE,CAAC;AAAA,IAC7C;AACA,UAAM,UAAU,KAAK,WAAW,iBAAiB,GAAG,KAAK,UAAU,OAAO,CAAC;AAAA,EAC7E;AAAA,EAEA,MAAM,eACJ,WACoC;AACpC,UAAM,WAAW,KAAK,WAAW,iBAAiB;AAClD,QAAI,CAAC,WAAW,QAAQ,EAAG,QAAO,oBAAI,IAAI;AAE1C,UAAM,MAAM,MAAM,SAAS,UAAU,OAAO;AAC5C,UAAM,UAAkD,KAAK,MAAM,GAAG;AACtE,UAAM,MAAM,oBAAI,IAA0B;AAC1C,eAAW,SAAS,SAAS;AAC3B,UAAI,IAAI,MAAM,KAAK,IAAI,aAAa,MAAM,IAAI,CAAC;AAAA,IACjD;AACA,WAAO;AAAA,EACT;AACF;;;AC5eA,OAAO,WAAW;AAClB,SAAS,mBAAmB;AAC5B,SAAS,qBAAqB;AAE9B,SAAS,YAAAA,WAAU,aAAAC,kBAAiB;AACpC,SAAS,QAAAC,aAAY;AACrB,SAAS,cAAAC,mBAAkB;AAQ3B,IAAM,oBAAoB;AAEnB,IAAM,eAAN,MAAmB;AAAA,EAChB;AAAA,EAER,cAAc;AACZ,SAAK,QAAQ,IAAI,MAAM,EAAE,MAAM,YAAY,OAAO,MAAM,CAAC;AAAA,EAC3D;AAAA,EAEA,mBAAmB,WAAoC;AACrD,SAAK,MAAM,MAAM;AAGjB,eAAW,OAAO,WAAW;AAC3B,WAAK,MAAM,QAAQ,IAAI,MAAM;AAAA,QAC3B,OAAO,IAAI;AAAA,QACX,MAAM,IAAI;AAAA,QACV,GAAI,IAAI,WAAW,UAAa,EAAE,QAAQ,IAAI,OAAO;AAAA,QACrD,GAAI,IAAI,iBAAiB,UAAa,EAAE,cAAc,IAAI,aAAa;AAAA,MACzE,CAAC;AAAA,IACH;AAEA,UAAM,aAAa,oBAAI,IAAoB;AAC3C,eAAW,OAAO,WAAW;AAC3B,YAAM,YAAY,IAAI,KAAK,QAAQ,SAAS,EAAE;AAC9C,YAAM,WAAW,UAAU,MAAM,GAAG,EAAE,IAAI;AAC1C,iBAAW,IAAI,SAAS,YAAY,GAAG,IAAI,IAAI;AAC/C,iBAAW,IAAI,UAAU,YAAY,GAAG,IAAI,IAAI;AAAA,IAClD;AAGA,eAAW,OAAO,WAAW;AAC3B,iBAAW,QAAQ,IAAI,WAAW;AAChC,cAAM,SAAS,WAAW,IAAI,KAAK,YAAY,CAAC;AAChD,YAAI,UAAU,WAAW,IAAI,QAAQ,CAAC,KAAK,MAAM,QAAQ,IAAI,MAAM,MAAM,GAAG;AAC1E,eAAK,MAAM,QAAQ,IAAI,MAAM,QAAQ,EAAE,MAAM,YAAY,QAAQ,EAAI,CAAC;AAAA,QACxE;AAAA,MACF;AAAA,IACF;AAIA,eAAW,OAAO,WAAW;AAC3B,UAAI,CAAC,IAAI,aAAa,OAAQ;AAC9B,iBAAW,OAAO,IAAI,aAAa;AAEjC,cAAM,aAAa,IAAI,QAAQ,SAAS,EAAE,EAAE,YAAY;AACxD,cAAM,OAAO,WAAW,MAAM,GAAG,EAAE,IAAI;AACvC,cAAM,SAAS,WAAW,IAAI,UAAU,KAAK,WAAW,IAAI,IAAI;AAChE,YAAI,UAAU,WAAW,IAAI,QAAQ,CAAC,KAAK,MAAM,QAAQ,IAAI,MAAM,MAAM,GAAG;AAC1E,eAAK,MAAM,QAAQ,IAAI,MAAM,QAAQ,EAAE,MAAM,WAAW,QAAQ,EAAI,CAAC;AAAA,QACvE;AAAA,MACF;AAAA,IACF;AAKA,UAAM,aAAa,oBAAI,IAAsB;AAC7C,eAAW,OAAO,WAAW;AAC3B,iBAAW,OAAO,IAAI,MAAM;AAC1B,cAAM,QAAQ,WAAW,IAAI,GAAG,KAAK,CAAC;AACtC,cAAM,KAAK,IAAI,IAAI;AACnB,mBAAW,IAAI,KAAK,KAAK;AAAA,MAC3B;AAAA,IACF;AAEA,eAAW,CAAC,EAAE,KAAK,KAAK,YAAY;AAClC,UAAI,MAAM,UAAU,kBAAmB;AACvC,eAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,iBAAS,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACzC,cAAI,CAAC,KAAK,MAAM,QAAQ,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,GAAG;AAC3C,iBAAK,MAAM,QAAQ,MAAM,CAAC,GAAG,MAAM,CAAC,GAAG,EAAE,MAAM,OAAO,QAAQ,IAAI,CAAC;AAAA,UACrE;AACA,cAAI,CAAC,KAAK,MAAM,QAAQ,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,GAAG;AAC3C,iBAAK,MAAM,QAAQ,MAAM,CAAC,GAAG,MAAM,CAAC,GAAG,EAAE,MAAM,OAAO,QAAQ,IAAI,CAAC;AAAA,UACrE;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAAA,EAEA,UAAU,UAA+B;AACvC,QAAI,CAAC,KAAK,MAAM,QAAQ,QAAQ,EAAG,QAAO,CAAC;AAC3C,WAAO,KAAK,MAAM,YAAY,QAAQ,EAAE,IAAI,CAAC,MAAM,KAAK,gBAAgB,CAAC,CAAC;AAAA,EAC5E;AAAA,EAEA,aAAa,UAA+B;AAC1C,QAAI,CAAC,KAAK,MAAM,QAAQ,QAAQ,EAAG,QAAO,CAAC;AAC3C,WAAO,KAAK,MAAM,aAAa,QAAQ,EAAE,IAAI,CAAC,MAAM,KAAK,gBAAgB,CAAC,CAAC;AAAA,EAC7E;AAAA,EAEA,SAAS,MAAc,IAA6B;AAClD,QAAI,CAAC,KAAK,MAAM,QAAQ,IAAI,KAAK,CAAC,KAAK,MAAM,QAAQ,EAAE,EAAG,QAAO;AACjE,UAAM,OAAO,cAAc,KAAK,OAAO,MAAM,EAAE;AAC/C,WAAO;AAAA,EACT;AAAA,EAEA,YAAY,SAAiB,WAAW,GAAG,QAAQ,IAAiB;AAClE,UAAM,KAAK,QAAQ,YAAY;AAE/B,UAAM,gBAAgB,IAAI;AAAA,MACxB,KAAK,MAAM,MAAM,EAAE,OAAO,CAAC,MAAM;AAC/B,cAAM,QAAQ,KAAK,MAAM,kBAAkB,CAAC;AAC5C,cAAM,QAAQ,OAAO,MAAM,UAAU,WAAW,MAAM,QAAQ,OAAO,MAAM,SAAS,EAAE;AACtF,eACE,EAAE,YAAY,EAAE,SAAS,EAAE,KAC3B,MAAM,YAAY,EAAE,SAAS,EAAE,KAC9B,MAAM,MAA+B,KAAK,CAAC,MAAM,EAAE,YAAY,EAAE,SAAS,EAAE,CAAC;AAAA,MAElF,CAAC;AAAA,IACH;AAEA,UAAM,UAAU,oBAAI,IAAY;AAChC,eAAW,SAAS,eAAe;AACjC,kBAAY,KAAK,OAAO,OAAO,CAAC,OAAO,OAAO,UAAU;AACtD,YAAI,QAAQ,QAAQ,QAAQ,EAAG,QAAO;AACtC,gBAAQ,IAAI,KAAK;AACjB,eAAO,SAAS;AAAA,MAClB,CAAC;AAAA,IACH;AAIA,UAAM,SAAS,CAAC,GAAG,OAAO,EAAE,KAAK,CAAC,GAAG,MAAM;AACzC,YAAM,SAAS,KAAK,MAAM,kBAAkB,CAAC;AAC7C,YAAM,SAAS,KAAK,MAAM,kBAAkB,CAAC;AAC7C,YAAM,cAAc,OAAO,WAAW,gBAAgB,OAAO,WAAW,aAAa,IAAI;AACzF,YAAM,cAAc,OAAO,WAAW,gBAAgB,OAAO,WAAW,aAAa,IAAI;AACzF,UAAI,gBAAgB,YAAa,QAAO,cAAc;AACtD,YAAM,UAAU,cAAc,IAAI,CAAC,IAAI,IAAI;AAC3C,YAAM,UAAU,cAAc,IAAI,CAAC,IAAI,IAAI;AAC3C,UAAI,YAAY,QAAS,QAAO,UAAU;AAC1C,YAAM,YAAa,OAAO,gBAAuC;AACjE,YAAM,YAAa,OAAO,gBAAuC;AACjE,UAAI,cAAc,UAAW,QAAO,YAAY;AAChD,aAAO,KAAK,MAAM,OAAO,CAAC,IAAI,KAAK,MAAM,OAAO,CAAC;AAAA,IACnD,CAAC;AAED,WAAO,OAAO,MAAM,GAAG,KAAK,EAAE,IAAI,CAAC,MAAM,KAAK,gBAAgB,CAAC,CAAC;AAAA,EAClE;AAAA,EAEA,aAAyB;AACvB,UAAM,QAAQ,KAAK,MAAM;AACzB,UAAM,QAAQ,KAAK,MAAM;AACzB,UAAM,UAAU,KAAK,MAAM,MAAM,EAAE;AAAA,MACjC,CAAC,MAAM,KAAK,MAAM,OAAO,CAAC,MAAM;AAAA,IAClC;AAEA,UAAM,cAAc,KAAK,MAAM,MAAM,EAAE,IAAI,CAAC,OAAO;AAAA,MACjD,MAAM;AAAA,MACN,aAAa,KAAK,MAAM,OAAO,CAAC;AAAA,IAClC,EAAE;AACF,gBAAY,KAAK,CAAC,GAAG,MAAM,EAAE,cAAc,EAAE,WAAW;AAExD,UAAM,mBAAmB,SAAS,QAAQ;AAC1C,UAAM,UAAU,mBAAmB,IAAI,QAAQ,mBAAmB;AAElE,WAAO;AAAA,MACL,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,aAAa,QAAQ;AAAA,MACrB,eAAe,YAAY,MAAM,GAAG,EAAE;AAAA,MACtC;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,gBAAgB,UAA6B;AACnD,UAAM,QAAQ,KAAK,MAAM,kBAAkB,QAAQ;AACnD,WAAO;AAAA,MACL,MAAM;AAAA,MACN,OAAO,OAAO,MAAM,UAAU,WAAW,MAAM,QAAQ;AAAA,MACvD,MAAO,MAAM,QAAqB,CAAC;AAAA,MACnC,WAAW,KAAK,MAAM,UAAU,QAAQ;AAAA,MACxC,eAAe,KAAK,MAAM,SAAS,QAAQ;AAAA,MAC3C,GAAI,MAAM,iBAAiB,UAAa,EAAE,cAAc,MAAM,aAAuB;AAAA,MACrF,GAAI,MAAM,WAAW,UAAa,EAAE,QAAQ,MAAM,OAAiB;AAAA,IACrE;AAAA,EACF;AAAA,EAEA,MAAM,KAAK,WAAkC;AAC3C,UAAM,OAAO,KAAK,MAAM,OAAO;AAC/B,UAAMF,WAAUC,MAAK,WAAW,YAAY,GAAG,KAAK,UAAU,IAAI,CAAC;AAAA,EACrE;AAAA,EAEA,MAAM,KAAK,WAAqC;AAC9C,UAAM,WAAWA,MAAK,WAAW,YAAY;AAC7C,QAAI,CAACC,YAAW,QAAQ,EAAG,QAAO;AAElC,UAAM,MAAM,MAAMH,UAAS,UAAU,OAAO;AAC5C,UAAM,OAAO,KAAK,MAAM,GAAG;AAC3B,SAAK,MAAM,OAAO,IAAI;AACtB,WAAO;AAAA,EACT;AAAA,EAEA,WAAkB;AAChB,WAAO,KAAK;AAAA,EACd;AACF;;;ACpNA,OAAO,aAAa;AAEpB,SAAS,YAAAI,WAAU,aAAAC,kBAAiB;AACpC,SAAS,QAAAC,aAAY;AACrB,SAAS,cAAAC,mBAAkB;AAH3B,IAAM,EAAE,gBAAgB,IAAI;AAYrB,IAAM,cAAN,MAAkB;AAAA,EACf,QAAqD;AAAA,EACrD;AAAA,EACA,YAAyB,CAAC;AAAA,EAElC,YAAY,YAAoB;AAC9B,SAAK,aAAa;AAAA,EACpB;AAAA,EAEA,MACE,YACA,MACM;AACN,QAAI,WAAW,WAAW,GAAG;AAC3B,WAAK,QAAQ;AACb,WAAK,YAAY,CAAC;AAClB;AAAA,IACF;AAEA,SAAK,QAAQ,IAAI,gBAAgB,UAAU,KAAK,UAAU;AAC1D,SAAK,MAAM,UAAU,WAAW,MAAM;AAEtC,aAAS,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;AAC1C,WAAK,MAAM,SAAS,MAAM,KAAK,WAAW,CAAC,CAAC,GAAG,CAAC;AAAA,IAClD;AAEA,SAAK,YAAY;AAAA,EACnB;AAAA,EAEA,OAAO,gBAA8B,IAAY,IAAoB;AACnE,QAAI,CAAC,KAAK,SAAS,KAAK,UAAU,WAAW,EAAG,QAAO,CAAC;AAExD,UAAM,aAAa,KAAK,IAAI,GAAG,KAAK,UAAU,MAAM;AACpD,UAAM,SAAS,KAAK,MAAM,UAAU,MAAM,KAAK,cAAc,GAAG,UAAU;AAE1E,UAAM,OAAO,oBAAI,IAAY;AAC7B,UAAM,UAA0B,CAAC;AAEjC,aAAS,IAAI,GAAG,IAAI,OAAO,UAAU,QAAQ,KAAK;AAChD,YAAM,MAAM,OAAO,UAAU,CAAC;AAC9B,YAAM,OAAO,KAAK,UAAU,GAAG;AAC/B,UAAI,CAAC,QAAQ,KAAK,IAAI,KAAK,OAAO,EAAG;AACrC,WAAK,IAAI,KAAK,OAAO;AAErB,cAAQ,KAAK;AAAA,QACX,MAAM,KAAK;AAAA,QACX,OAAO,KAAK;AAAA,QACZ,OAAO,IAAI,OAAO,UAAU,CAAC;AAAA,QAC7B,SAAS,KAAK,KAAK,MAAM,GAAG,GAAG;AAAA,QAC/B,cAAc,KAAK;AAAA,MACrB,CAAC;AAAA,IACH;AAEA,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,KAAK,WAAkC;AAC3C,QAAI,CAAC,KAAK,MAAO;AAEjB,SAAK,MAAM,eAAeD,MAAK,WAAW,UAAU,CAAC;AACrD,UAAMD;AAAA,MACJC,MAAK,WAAW,gBAAgB;AAAA,MAChC,KAAK,UAAU,KAAK,SAAS;AAAA,IAC/B;AAAA,EACF;AAAA,EAEA,MAAM,KAAK,WAAqC;AAC9C,UAAM,WAAWA,MAAK,WAAW,UAAU;AAC3C,UAAM,WAAWA,MAAK,WAAW,gBAAgB;AAEjD,QAAI,CAACC,YAAW,QAAQ,KAAK,CAACA,YAAW,QAAQ,EAAG,QAAO;AAE3D,UAAM,MAAM,MAAMH,UAAS,UAAU,OAAO;AAC5C,SAAK,YAAY,KAAK,MAAM,GAAG;AAE/B,SAAK,QAAQ,IAAI,gBAAgB,UAAU,KAAK,UAAU;AAC1D,SAAK,MAAM,UAAU,KAAK,UAAU,MAAM;AAC1C,SAAK,MAAM,cAAc,QAAQ;AAEjC,WAAO;AAAA,EACT;AAAA,EAEA,eAA4B;AAC1B,WAAO,KAAK;AAAA,EACd;AACF;;;ACjGA,SAAS,iBAAiB;AAEnB,IAAM,aAAN,MAAiB;AAAA,EACd,YAA+B,CAAC;AAAA,EAExC,aAAa,WAAoC;AAC/C,SAAK,YAAY;AAAA,EACnB;AAAA,EAEA,OAAO,SAA4C;AACjD,UAAM,EAAE,SAAS,OAAO,eAAe,UAAU,WAAW,QAAQ,GAAG,IAAI;AAE3E,QAAI;AAEJ,QAAI,OAAO;AACT,YAAM,QAAQ,gBAAgB,MAAM;AACpC,YAAM,KAAK,IAAI,OAAO,SAAS,KAAK;AACpC,gBAAU,CAAC,SAAS;AAClB,WAAG,YAAY;AACf,cAAM,IAAI,GAAG,KAAK,IAAI;AACtB,eAAO,EAAE,SAAS,CAAC,CAAC,GAAG,OAAO,GAAG,SAAS,GAAG;AAAA,MAC/C;AAAA,IACF,OAAO;AACL,YAAM,SAAS,gBAAgB,UAAU,QAAQ,YAAY;AAC7D,gBAAU,CAAC,SAAS;AAClB,cAAM,WAAW,gBAAgB,OAAO,KAAK,YAAY;AACzD,cAAM,MAAM,SAAS,QAAQ,MAAM;AACnC,eAAO,EAAE,SAAS,OAAO,GAAG,OAAO,IAAI;AAAA,MACzC;AAAA,IACF;AAEA,UAAM,UAA0B,CAAC;AAEjC,eAAW,OAAO,KAAK,WAAW;AAEhC,UAAI,YAAY,CAAC,UAAU,IAAI,MAAM,QAAQ,EAAG;AAGhD,UAAI,WAAW,QAAQ;AACrB,cAAM,SAAS,UAAU,KAAK,CAAC,MAAM,IAAI,KAAK,SAAS,CAAC,CAAC;AACzD,YAAI,CAAC,OAAQ;AAAA,MACf;AAEA,YAAM,EAAE,SAAS,MAAM,IAAI,QAAQ,IAAI,OAAO;AAC9C,UAAI,CAAC,QAAS;AAEd,YAAM,eAAe,KAAK,IAAI,GAAG,QAAQ,EAAE;AAC3C,YAAM,aAAa,KAAK,IAAI,IAAI,QAAQ,QAAQ,QAAQ,GAAG;AAC3D,YAAM,UAAU,IAAI,QAAQ,MAAM,cAAc,UAAU,EAAE,KAAK;AAEjE,cAAQ,KAAK;AAAA,QACX,MAAM,IAAI;AAAA,QACV,OAAO,IAAI;AAAA,QACX,OAAO;AAAA,QACP,UAAU,eAAe,IAAI,QAAQ,MAAM,WAAW,aAAa,IAAI,QAAQ,SAAS,QAAQ;AAAA,MAClG,CAAC;AAED,UAAI,QAAQ,UAAU,MAAO;AAAA,IAC/B;AAEA,WAAO;AAAA,EACT;AACF;","names":["readFile","writeFile","join","existsSync","readFile","writeFile","join","existsSync"]}

package/dist/chunk-5F5ZIAW2.js

@@ -1,79 +0,0 @@
-// src/core/failure-eval.ts
-var key = (c, p) => `${c}\0${p}`;
-function evaluateFailures(results, judgments, opts = {}) {
-  const target = opts.targetPrecision ?? 0.95;
-  const minJudged = opts.minJudged ?? 20;
-  const thresholds = opts.thresholds ?? Array.from({ length: 17 }, (_, i) => Math.round((0.2 + i * 0.05) * 100) / 100);
-  const label = /* @__PURE__ */ new Map();
-  const relevantFor = /* @__PURE__ */ new Map();
-  for (const j of judgments) {
-    label.set(key(j.case, j.path), (label.get(key(j.case, j.path)) ?? false) || j.relevant);
-    if (j.relevant) {
-      if (!relevantFor.has(j.case)) relevantFor.set(j.case, /* @__PURE__ */ new Set());
-      relevantFor.get(j.case).add(j.path);
-    }
-  }
-  let answerable = 0, rank1 = 0, recallAtK = 0;
-  const unjudged = [];
-  for (const r of results) {
-    const relevant = relevantFor.get(r.case);
-    if (relevant && relevant.size > 0) {
-      answerable++;
-      if (r.hits[0] && relevant.has(r.hits[0].path)) rank1++;
-      if (r.hits.some((h) => relevant.has(h.path))) recallAtK++;
-    }
-    const top = r.hits[0];
-    if (top && !label.has(key(r.case, top.path))) unjudged.push({ case: r.case, path: top.path, score: top.score });
-  }
-  const rows = thresholds.map((threshold) => {
-    let injected = 0, judged = 0, correct = 0;
-    for (const r of results) {
-      const top = r.hits[0];
-      if (!top || top.score < threshold) continue;
-      injected++;
-      const l = label.get(key(r.case, top.path));
-      if (l === void 0) continue;
-      judged++;
-      if (l) correct++;
-    }
-    return {
-      threshold,
-      injected,
-      judged,
-      correct,
-      precision: judged > 0 ? correct / judged : null,
-      coverage: results.length > 0 ? injected / results.length : 0
-    };
-  });
-  const meets = (row) => row.precision !== null && row.precision >= target;
-  const recommended = rows.find(
-    (row, i) => row.judged >= minJudged && meets(row) && rows.slice(i + 1).every((r) => r.judged < minJudged || meets(r))
-  )?.threshold ?? null;
-  return { cases: results.length, answerable, rank1, recallAtK, rows, recommended, unjudged };
-}
-function formatFailureEval(report, label = "") {
-  const pct = (n) => n === null ? "  -  " : `${(n * 100).toFixed(1)}%`.padStart(6);
-  const lines = [
-    `Failure-lookup eval${label ? ` (${label})` : ""} \u2014 ${report.cases} cases, ${report.answerable} with a known answer`,
-    `  answer ranked first: ${report.rank1}/${report.answerable}   answer anywhere in hits: ${report.recallAtK}/${report.answerable}`,
-    "",
-    "  threshold  injected  judged  correct  precision  coverage"
-  ];
-  for (const r of report.rows) {
-    lines.push(
-      `  ${r.threshold.toFixed(2).padStart(9)}  ${String(r.injected).padStart(8)}  ${String(r.judged).padStart(6)}  ${String(r.correct).padStart(7)}  ${pct(r.precision).padStart(9)}  ${pct(r.coverage).padStart(8)}`
-    );
-  }
-  lines.push("");
-  lines.push(
-    report.recommended === null ? "  recommended threshold: none reaches the target precision with enough judged injections" : `  recommended threshold: ${report.recommended.toFixed(2)}`
-  );
-  lines.push(`  unjudged top hits: ${report.unjudged.length}`);
-  return lines.join("\n");
-}
-
-export {
-  evaluateFailures,
-  formatFailureEval
-};
-//# sourceMappingURL=chunk-5F5ZIAW2.js.map

package/dist/chunk-5F5ZIAW2.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/core/failure-eval.ts"],"sourcesContent":["/**\n * Failure-lookup eval (v1.6): given failed tool calls, the hits a search returned for each, and human\n * (or judge) relevance labels, measure how well the top hit answers the failure and pick the score\n * threshold above which injecting it is right often enough. Pure — no I/O, no search.\n */\n\nexport interface FailureCase {\n  id: string;\n  /** notes folder this failure's repo searches (its docs) */\n  notes: string;\n  error: string;\n  command: string;\n}\n\nexport interface Judgment {\n  case: string;\n  path: string;\n  relevant: boolean;\n}\n\nexport interface CaseResult {\n  case: string;\n  hits: { path: string; score: number }[];\n}\n\nexport interface ThresholdRow {\n  threshold: number;\n  /** cases whose top hit scored >= threshold (would be injected) */\n  injected: number;\n  /** injected cases whose top hit has a label */\n  judged: number;\n  correct: number;\n  /** correct / judged; null when nothing judged */\n  precision: number | null;\n  /** injected / all cases */\n  coverage: number;\n}\n\nexport interface FailureEvalReport {\n  cases: number;\n  /** cases with at least one relevant path among their labels */\n  answerable: number;\n  /** answerable cases whose top hit is relevant */\n  rank1: number;\n  /** answerable cases with a relevant path anywhere in the hits */\n  recallAtK: number;\n  rows: ThresholdRow[];\n  /** lowest threshold reaching the target precision on >= minJudged judged injections; null if none */\n  recommended: number | null;\n  /** top hits nobody has labelled yet: label these, then re-run */\n  unjudged: { case: string; path: string; score: number }[];\n}\n\nconst key = (c: string, p: string) => `${c}\\u0000${p}`;\n\nexport function evaluateFailures(\n  results: CaseResult[],\n  judgments: Judgment[],\n  opts: { targetPrecision?: number; minJudged?: number; thresholds?: number[] } = {}\n): FailureEvalReport {\n  const target = opts.targetPrecision ?? 0.95;\n  const minJudged = opts.minJudged ?? 20;\n  const thresholds = opts.thresholds ?? Array.from({ length: 17 }, (_, i) => Math.round((0.2 + i * 0.05) * 100) / 100);\n\n  const label = new Map<string, boolean>();\n  const relevantFor = new Map<string, Set<string>>();\n  for (const j of judgments) {\n    // a positive label wins over a negative one for the same pair\n    label.set(key(j.case, j.path), (label.get(key(j.case, j.path)) ?? false) || j.relevant);\n    if (j.relevant) {\n      if (!relevantFor.has(j.case)) relevantFor.set(j.case, new Set());\n      relevantFor.get(j.case)!.add(j.path);\n    }\n  }\n\n  let answerable = 0, rank1 = 0, recallAtK = 0;\n  const unjudged: FailureEvalReport[\"unjudged\"] = [];\n  for (const r of results) {\n    const relevant = relevantFor.get(r.case);\n    if (relevant && relevant.size > 0) {\n      answerable++;\n      if (r.hits[0] && relevant.has(r.hits[0].path)) rank1++;\n      if (r.hits.some((h) => relevant.has(h.path))) recallAtK++;\n    }\n    const top = r.hits[0];\n    if (top && !label.has(key(r.case, top.path))) unjudged.push({ case: r.case, path: top.path, score: top.score });\n  }\n\n  const rows: ThresholdRow[] = thresholds.map((threshold) => {\n    let injected = 0, judged = 0, correct = 0;\n    for (const r of results) {\n      const top = r.hits[0];\n      if (!top || top.score < threshold) continue;\n      injected++;\n      const l = label.get(key(r.case, top.path));\n      if (l === undefined) continue;\n      judged++;\n      if (l) correct++;\n    }\n    return {\n      threshold,\n      injected,\n      judged,\n      correct,\n      precision: judged > 0 ? correct / judged : null,\n      coverage: results.length > 0 ? injected / results.length : 0,\n    };\n  });\n\n  // lowest threshold where this row and every higher row with enough judged injections reach the target\n  const meets = (row: ThresholdRow) => row.precision !== null && row.precision >= target;\n  const recommended =\n    rows.find(\n      (row, i) => row.judged >= minJudged && meets(row) && rows.slice(i + 1).every((r) => r.judged < minJudged || meets(r))\n    )?.threshold ?? null;\n\n  return { cases: results.length, answerable, rank1, recallAtK, rows, recommended, unjudged };\n}\n\nexport function formatFailureEval(report: FailureEvalReport, label = \"\"): string {\n  const pct = (n: number | null) => (n === null ? \"  -  \" : `${(n * 100).toFixed(1)}%`.padStart(6));\n  const lines = [\n    `Failure-lookup eval${label ? ` (${label})` : \"\"} — ${report.cases} cases, ${report.answerable} with a known answer`,\n    `  answer ranked first: ${report.rank1}/${report.answerable}   answer anywhere in hits: ${report.recallAtK}/${report.answerable}`,\n    \"\",\n    \"  threshold  injected  judged  correct  precision  coverage\",\n  ];\n  for (const r of report.rows) {\n    lines.push(\n      `  ${r.threshold.toFixed(2).padStart(9)}  ${String(r.injected).padStart(8)}  ${String(r.judged).padStart(6)}  ${String(r.correct).padStart(7)}  ${pct(r.precision).padStart(9)}  ${pct(r.coverage).padStart(8)}`\n    );\n  }\n  lines.push(\"\");\n  lines.push(\n    report.recommended === null\n      ? \"  recommended threshold: none reaches the target precision with enough judged injections\"\n      : `  recommended threshold: ${report.recommended.toFixed(2)}`\n  );\n  lines.push(`  unjudged top hits: ${report.unjudged.length}`);\n  return lines.join(\"\\n\");\n}\n"],"mappings":";AAqDA,IAAM,MAAM,CAAC,GAAW,MAAc,GAAG,CAAC,KAAS,CAAC;AAE7C,SAAS,iBACd,SACA,WACA,OAAgF,CAAC,GAC9D;AACnB,QAAM,SAAS,KAAK,mBAAmB;AACvC,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,aAAa,KAAK,cAAc,MAAM,KAAK,EAAE,QAAQ,GAAG,GAAG,CAAC,GAAG,MAAM,KAAK,OAAO,MAAM,IAAI,QAAQ,GAAG,IAAI,GAAG;AAEnH,QAAM,QAAQ,oBAAI,IAAqB;AACvC,QAAM,cAAc,oBAAI,IAAyB;AACjD,aAAW,KAAK,WAAW;AAEzB,UAAM,IAAI,IAAI,EAAE,MAAM,EAAE,IAAI,IAAI,MAAM,IAAI,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,KAAK,UAAU,EAAE,QAAQ;AACtF,QAAI,EAAE,UAAU;AACd,UAAI,CAAC,YAAY,IAAI,EAAE,IAAI,EAAG,aAAY,IAAI,EAAE,MAAM,oBAAI,IAAI,CAAC;AAC/D,kBAAY,IAAI,EAAE,IAAI,EAAG,IAAI,EAAE,IAAI;AAAA,IACrC;AAAA,EACF;AAEA,MAAI,aAAa,GAAG,QAAQ,GAAG,YAAY;AAC3C,QAAM,WAA0C,CAAC;AACjD,aAAW,KAAK,SAAS;AACvB,UAAM,WAAW,YAAY,IAAI,EAAE,IAAI;AACvC,QAAI,YAAY,SAAS,OAAO,GAAG;AACjC;AACA,UAAI,EAAE,KAAK,CAAC,KAAK,SAAS,IAAI,EAAE,KAAK,CAAC,EAAE,IAAI,EAAG;AAC/C,UAAI,EAAE,KAAK,KAAK,CAAC,MAAM,SAAS,IAAI,EAAE,IAAI,CAAC,EAAG;AAAA,IAChD;AACA,UAAM,MAAM,EAAE,KAAK,CAAC;AACpB,QAAI,OAAO,CAAC,MAAM,IAAI,IAAI,EAAE,MAAM,IAAI,IAAI,CAAC,EAAG,UAAS,KAAK,EAAE,MAAM,EAAE,MAAM,MAAM,IAAI,MAAM,OAAO,IAAI,MAAM,CAAC;AAAA,EAChH;AAEA,QAAM,OAAuB,WAAW,IAAI,CAAC,cAAc;AACzD,QAAI,WAAW,GAAG,SAAS,GAAG,UAAU;AACxC,eAAW,KAAK,SAAS;AACvB,YAAM,MAAM,EAAE,KAAK,CAAC;AACpB,UAAI,CAAC,OAAO,IAAI,QAAQ,UAAW;AACnC;AACA,YAAM,IAAI,MAAM,IAAI,IAAI,EAAE,MAAM,IAAI,IAAI,CAAC;AACzC,UAAI,MAAM,OAAW;AACrB;AACA,UAAI,EAAG;AAAA,IACT;AACA,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,WAAW,SAAS,IAAI,UAAU,SAAS;AAAA,MAC3C,UAAU,QAAQ,SAAS,IAAI,WAAW,QAAQ,SAAS;AAAA,IAC7D;AAAA,EACF,CAAC;AAGD,QAAM,QAAQ,CAAC,QAAsB,IAAI,cAAc,QAAQ,IAAI,aAAa;AAChF,QAAM,cACJ,KAAK;AAAA,IACH,CAAC,KAAK,MAAM,IAAI,UAAU,aAAa,MAAM,GAAG,KAAK,KAAK,MAAM,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,SAAS,aAAa,MAAM,CAAC,CAAC;AAAA,EACtH,GAAG,aAAa;AAElB,SAAO,EAAE,OAAO,QAAQ,QAAQ,YAAY,OAAO,WAAW,MAAM,aAAa,SAAS;AAC5F;AAEO,SAAS,kBAAkB,QAA2B,QAAQ,IAAY;AAC/E,QAAM,MAAM,CAAC,MAAsB,MAAM,OAAO,UAAU,IAAI,IAAI,KAAK,QAAQ,CAAC,CAAC,IAAI,SAAS,CAAC;AAC/F,QAAM,QAAQ;AAAA,IACZ,sBAAsB,QAAQ,KAAK,KAAK,MAAM,EAAE,WAAM,OAAO,KAAK,WAAW,OAAO,UAAU;AAAA,IAC9F,0BAA0B,OAAO,KAAK,IAAI,OAAO,UAAU,+BAA+B,OAAO,SAAS,IAAI,OAAO,UAAU;AAAA,IAC/H;AAAA,IACA;AAAA,EACF;AACA,aAAW,KAAK,OAAO,MAAM;AAC3B,UAAM;AAAA,MACJ,KAAK,EAAE,UAAU,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,KAAK,OAAO,EAAE,QAAQ,EAAE,SAAS,CAAC,CAAC,KAAK,OAAO,EAAE,MAAM,EAAE,SAAS,CAAC,CAAC,KAAK,OAAO,EAAE,OAAO,EAAE,SAAS,CAAC,CAAC,KAAK,IAAI,EAAE,SAAS,EAAE,SAAS,CAAC,CAAC,KAAK,IAAI,EAAE,QAAQ,EAAE,SAAS,CAAC,CAAC;AAAA,IAChN;AAAA,EACF;AACA,QAAM,KAAK,EAAE;AACb,QAAM;AAAA,IACJ,OAAO,gBAAgB,OACnB,6FACA,4BAA4B,OAAO,YAAY,QAAQ,CAAC,CAAC;AAAA,EAC/D;AACA,QAAM,KAAK,wBAAwB,OAAO,SAAS,MAAM,EAAE;AAC3D,SAAO,MAAM,KAAK,IAAI;AACxB;","names":[]}