Skip to content

armgit / xerj

Public Imported from GitHub
0

AI search engine and MCP server for bots and agents. One Rust binary indexes code, docs, logs, JSONL/CSV and PDFs for BM25 full-text, vector and hybrid search, code search without grep, agent memory and RAG. Elasticsearch-compatible API; Corpus Hub serves hundreds of reference memory corpora.

39.3 KB
Raw Blame
<!DOCTYPE html>
<html lang="en" data-theme="day">
<head>
<!-- ga4:begin — generated by scripts/seo/seo_head.py · do not hand-edit -->
<!-- Google tag (gtag.js) -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-Z77ETCDJYL"></script>
<script>
  window.dataLayer = window.dataLayer || [];
  function gtag(){dataLayer.push(arguments);}
  gtag('js', new Date());

  gtag('config', 'G-Z77ETCDJYL');
</script>
<!-- ga4:end -->
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>XERJ · Docs · Recipe · Vector quantization</title>
<meta name="description" content="Opt a dense_vector field into scalar8 quantization: kNN scored from 1-byte-per-dimension codes at recall@10 = 0.998, originals still returned in _source.">
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Big+Shoulders+Display:wght@400;700;800;900&family=Inter:wght@400;500;600;700&family=IBM+Plex+Sans:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500;700&display=swap" rel="stylesheet">
<link rel="icon" href="data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'><rect width='100' height='100' fill='%230b0b0d'/><text x='50' y='68' font-size='60' font-family='Big Shoulders Display,sans-serif' font-weight='900' fill='%23ffc400' text-anchor='middle'>X</text></svg>">
<script src="/theme.js"></script>
<link rel="stylesheet" href="/style.css?v=90f643e">
<!-- seo:begin — generated by scripts/seo/fix_heads.py · do not hand-edit -->
<link rel="canonical" href="https://xerj.org/docs/recipes/vector-quantization">
<meta property="og:site_name" content="XERJ.ai">
<meta property="og:type" content="article">
<meta property="og:title" content="XERJ · Docs · Recipe · Vector quantization">
<meta property="og:description" content="Opt a dense_vector field into scalar8 quantization: kNN scored from 1-byte-per-dimension codes at recall@10 = 0.998, originals still returned in _source.">
<meta property="og:url" content="https://xerj.org/docs/recipes/vector-quantization">
<meta property="og:image" content="https://xerj.org/og/xerj-card.png">
<meta property="og:image:width" content="1200">
<meta property="og:image:height" content="630">
<meta property="og:image:alt" content="XERJ.ai — the search engine built for AI agents">
<meta property="og:locale" content="en_US">
<meta name="twitter:card" content="summary_large_image">
<meta name="twitter:title" content="XERJ · Docs · Recipe · Vector quantization">
<meta name="twitter:description" content="Opt a dense_vector field into scalar8 quantization: kNN scored from 1-byte-per-dimension codes at recall@10 = 0.998, originals still returned in _source.">
<meta name="twitter:image" content="https://xerj.org/og/xerj-card.png">
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "@id": "https://xerj.org/#organization",
  "name": "XERJ",
  "alternateName": "XERJ.ai",
  "url": "https://xerj.org/",
  "logo": "https://xerj.org/brandbook/xerj-mark.svg",
  "description": "A new way for AI to search for data: one local Rust binary that indexes your code, docs, specs and logs, plus Corpus Hub packs, and returns cited answers.",
  "sameAs": [
    "https://github.com/xerj-org/xerj",
    "https://x.com/xerj_org"
  ]
}
</script>
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "WebSite",
  "@id": "https://xerj.org/#website",
  "name": "XERJ.ai",
  "url": "https://xerj.org/",
  "inLanguage": "en",
  "publisher": {
    "@id": "https://xerj.org/#organization"
  }
}
</script>
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "TechArticle",
  "headline": "XERJ · Docs · Recipe · Vector quantization",
  "description": "Opt a dense_vector field into scalar8 quantization: kNN scored from 1-byte-per-dimension codes at recall@10 = 0.998, originals still returned in _source.",
  "url": "https://xerj.org/docs/recipes/vector-quantization",
  "image": "https://xerj.org/og/xerj-card.png",
  "inLanguage": "en",
  "datePublished": "2026-07-08",
  "dateModified": "2026-09-08",
  "author": {
    "@id": "https://xerj.org/#organization"
  },
  "publisher": {
    "@id": "https://xerj.org/#organization"
  },
  "isPartOf": {
    "@id": "https://xerj.org/#website"
  },
  "mainEntityOfPage": {
    "@type": "WebPage",
    "@id": "https://xerj.org/docs/recipes/vector-quantization"
  }
}
</script>
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "BreadcrumbList",
  "itemListElement": [
    {
      "@type": "ListItem",
      "position": 1,
      "name": "Home",
      "item": "https://xerj.org/"
    },
    {
      "@type": "ListItem",
      "position": 2,
      "name": "Docs",
      "item": "https://xerj.org/docs/"
    },
    {
      "@type": "ListItem",
      "position": 3,
      "name": "Recipes",
      "item": "https://xerj.org/docs/recipes/"
    },
    {
      "@type": "ListItem",
      "position": 4,
      "name": "Vector quantization"
    }
  ]
}
</script>
<!-- seo:end -->
</head>
<body>


<nav class="nav" aria-label="Primary">
  <a href="/" class="brand">XERJ</a>
  <a href="/product">PRODUCT</a>
  <a href="/use-cases">USE CASES</a>
  <a href="/answers/">ANSWERS</a>
  <a href="/benchmarks/">BENCHMARKS</a>
  <a href="/blog/">BLOG</a>
  <span class="spacer"></span>
  <a href="/for-agents">FOR AGENTS</a>
  <a href="/docs/" class="active">DOCS</a>
  <a href="/playground">DASHBOARDS</a>
  <a href="https://hub.xerj.org" rel="external">HUB</a>
  <a class="nav-stars" href="https://github.com/xerj-org/xerj" rel="external" title="XERJ on GitHub" aria-label="XERJ on GitHub, 14,225 stars"><svg viewBox="0 0 16 16" width="14" height="14" aria-hidden="true"><path d="M8 0C3.58 0 0 3.58 0 8c0 3.54 2.29 6.53 5.47 7.59.4.07.55-.17.55-.38 0-.19-.01-.82-.01-1.49-2.01.37-2.53-.49-2.69-.94-.09-.23-.48-.94-.82-1.13-.28-.15-.68-.52-.01-.53.63-.01 1.08.58 1.23.82.72 1.21 1.87.87 2.33.66.07-.52.28-.87.51-1.07-1.78-.2-3.64-.89-3.64-3.95 0-.87.31-1.59.82-2.15-.08-.2-.36-1.02.08-2.12 0 0 .67-.21 2.2.82.64-.18 1.32-.27 2-.27.68 0 1.36.09 2 .27 1.53-1.04 2.2-.82 2.2-.82.44 1.1.16 1.92.08 2.12.51.56.82 1.27.82 2.15 0 3.07-1.87 3.75-3.65 3.95.29.25.54.73.54 1.48 0 1.07-.01 1.93-.01 2.2 0 .21.15.46.55.38A8.013 8.013 0 0016 8c0-4.42-3.58-8-8-8z"/></svg><span class="nav-stars-n" data-gh-stars="14225">★ 14.2k</span></a>
  <a href="https://x.com/xerj_org" rel="external" title="XERJ on X" style="display:inline-flex;align-items:center"><svg width="15" height="15" viewBox="0 0 24 24" fill="currentColor" aria-hidden="true"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z"/></svg></a>
  <a href="https://discord.com/invite/yetXFusXfs" rel="external" title="XERJ on Discord" style="display:inline-flex;align-items:center"><svg width="15" height="15" viewBox="0 0 24 24" fill="currentColor" aria-hidden="true"><path d="M20.317 4.37a19.79 19.79 0 00-4.885-1.515.074.074 0 00-.079.037c-.21.375-.444.864-.608 1.25a18.27 18.27 0 00-5.487 0 12.64 12.64 0 00-.617-1.25.077.077 0 00-.079-.037A19.74 19.74 0 003.677 4.37a.07.07 0 00-.032.027C.533 9.046-.32 13.58.099 18.057a.082.082 0 00.031.057 19.9 19.9 0 005.993 3.03.078.078 0 00.084-.028c.462-.63.874-1.295 1.226-1.994a.076.076 0 00-.041-.106 13.1 13.1 0 01-1.872-.892.077.077 0 01-.008-.128 10.2 10.2 0 00.372-.292.074.074 0 01.077-.01c3.928 1.793 8.18 1.793 12.062 0a.074.074 0 01.078.01c.12.098.246.198.373.292a.077.077 0 01-.006.127 12.3 12.3 0 01-1.873.892.077.077 0 00-.041.107c.36.698.772 1.362 1.225 1.993a.076.076 0 00.084.028 19.84 19.84 0 006.002-3.03.077.077 0 00.032-.054c.5-5.177-.838-9.674-3.549-13.66a.061.061 0 00-.031-.03zM8.02 15.33c-1.183 0-2.157-1.085-2.157-2.419 0-1.333.956-2.419 2.157-2.419 1.21 0 2.176 1.096 2.157 2.42 0 1.333-.956 2.418-2.157 2.418zm7.975 0c-1.183 0-2.157-1.085-2.157-2.419 0-1.333.955-2.419 2.157-2.419 1.21 0 2.176 1.096 2.157 2.42 0 1.333-.946 2.418-2.157 2.418z"/></svg></a>
  <span class="theme">
    <button type="button" data-theme="day" class="active">DAY</button>
    <span class="dash">·</span>
    <button type="button" data-theme="night">NIGHT</button>
  </span>
</nav>

<div class="docs-search" id="docsSearch">
  <input type="text" id="docsSearchInput" placeholder="Search docs · keys, flags, endpoints, playbooks…" aria-label="Search docs" autocomplete="off" spellcheck="false">
  <span class="hint">⌘ K</span>
  <div class="results" id="docsSearchResults" aria-live="polite"></div>
</div>

<div class="docs-shell">
<nav class="docs-nav" aria-label="Docs">  <div class="group">
    <span class="label">START</span>
    <a href="/docs/" class=""><span class="eyebrow">01</span>Overview</a>
      <a href="/docs/quickstart" class=""><span class="eyebrow">02</span>Quickstart</a>
      <a href="/docs/install" class=""><span class="eyebrow">03</span>Install</a>
      <a href="/docs/migration-from-es" class=""><span class="eyebrow">04</span>Migrate from ES</a>
  </div>
  <div class="group">
    <span class="label">REFERENCE</span>
    <a href="/docs/cli" class=""><span class="eyebrow">01</span>CLI</a>
      <a href="/docs/config" class=""><span class="eyebrow">02</span>Config TOML</a>
      <a href="/docs/env" class=""><span class="eyebrow">03</span>Environment</a>
      <a href="/docs/api-native" class=""><span class="eyebrow">04</span>Native REST API</a>
      <a href="/docs/api-es-compat" class=""><span class="eyebrow">05</span>ES-compat API</a>
  </div>
  <div class="group">
    <span class="label">DATA MODEL</span>
    <a href="/docs/queries" class=""><span class="eyebrow">01</span>Query types</a>
      <a href="/docs/aggregations" class=""><span class="eyebrow">02</span>Aggregations</a>
      <a href="/docs/analyzers" class=""><span class="eyebrow">03</span>Analyzers</a>
      <a href="/docs/vectors" class=""><span class="eyebrow">04</span>Vectors & kNN</a>
  </div>
  <div class="group">
    <span class="label">ENGINE</span>
    <a href="/docs/ingest" class=""><span class="eyebrow">01</span>Ingest pipelines</a>
      <a href="/docs/storage" class=""><span class="eyebrow">02</span>Storage & WAL</a>
      <a href="/docs/compression" class=""><span class="eyebrow">03</span>Compression</a>
      <a href="/docs/clustering" class=""><span class="eyebrow">04</span>Clustering</a>
      <a href="/docs/metrics" class=""><span class="eyebrow">05</span>Metrics</a>
      <a href="/docs/security" class=""><span class="eyebrow">06</span>Auth & TLS</a>
  </div>
  <div class="group">
    <span class="label">OPERATIONS</span>
    <a href="/docs/operations" class=""><span class="eyebrow">01</span>Running in production</a>
      <a href="/docs/backup-restore" class=""><span class="eyebrow">02</span>Backup & restore</a>
      <a href="/docs/upgrades" class=""><span class="eyebrow">03</span>Upgrades</a>
      <a href="/docs/troubleshooting" class=""><span class="eyebrow">04</span>Troubleshooting</a>
  </div>
  <div class="group">
    <span class="label">PLAYBOOKS</span>
    <a href="/docs/playbooks/siem" class=""><span class="eyebrow">01</span>SIEM</a>
      <a href="/docs/playbooks/log-analytics" class=""><span class="eyebrow">02</span>Log analytics</a>
      <a href="/docs/playbooks/vector-search" class=""><span class="eyebrow">03</span>Vector · RAG</a>
      <a href="/docs/playbooks/full-text" class=""><span class="eyebrow">04</span>Full-text search</a>
      <a href="/docs/playbooks/observability" class=""><span class="eyebrow">05</span>Observability</a>
  </div>
  <div class="group">
    <span class="label">RECIPES</span>
    <a href="/docs/recipes/" class=""><span class="eyebrow">00</span>All recipes</a>
    <a href="/docs/recipes/semantic-search-rag" class=""><span class="eyebrow">01</span>Semantic search &amp; RAG</a>
    <a href="/docs/recipes/agentic-memory" class=""><span class="eyebrow">02</span>Agent memory</a>
    <a href="/docs/recipes/passage-retrieval" class=""><span class="eyebrow">03</span>Passage retrieval</a>
    <a href="/docs/recipes/vector-search-knn" class=""><span class="eyebrow">04</span>Vector search &middot; kNN</a>
    <a href="/docs/recipes/vector-quantization" class="active"><span class="eyebrow">05</span>Vector quantization</a>
    <a href="/docs/recipes/hybrid-search" class=""><span class="eyebrow">06</span>Hybrid search</a>
    <a href="/docs/recipes/log-analytics" class=""><span class="eyebrow">07</span>Log analytics</a>
    <a href="/docs/recipes/anomaly-detection" class=""><span class="eyebrow">08</span>Anomaly detection</a>
    <a href="/docs/recipes/continuous-anomaly-datafeeds" class=""><span class="eyebrow">09</span>Anomaly datafeeds</a>
    <a href="/docs/recipes/migrate-from-elasticsearch" class=""><span class="eyebrow">10</span>Migrate from ES</a>
  </div>
</nav>

<main class="docs-content">
  <!-- Print-only branded header (hidden on screen, visible in PDF) -->
  <div class="print-header" style="display:none;">XERJ</div>

  <div class="breadcrumb"><a href="/docs/">DOCS</a><span class="sep">·</span><a href="/docs/recipes/">RECIPES</a><span class="sep">·</span>VECTOR QUANTIZATION
    <button class="pdf-btn" onclick="window.print()" title="Save this page as PDF">
      <svg viewBox="0 0 24 24"><path d="M12 3v12M12 15l-4-4M12 15l4-4"/><path d="M4 17v2a2 2 0 002 2h12a2 2 0 002-2v-2"/></svg>
      PDF
    </button>
  </div>
  <div class="kicker"><span>RECIPE · 05</span></div>
  <h1>Vector quantization</h1>

  <p><strong>Read this first.</strong> <code class="inline">scalar8</code> on XERJ today changes <strong>precision, not memory</strong>. Scores are computed from 1-byte-per-dimension codes, so the field has the recall profile of int8 — but the serving path still reads the full-precision vector out of <code class="inline">_source</code> and quantizes it per query, so nothing gets smaller in RAM. The ingest-time code array that would make it a memory win is tracked in <a href="https://github.com/xerj-org/xerj/issues/392">#392</a>. If you came here for a smaller working set, that issue is the one to watch, not this recipe.</p>

  <h2>The problem</h2>
  <p>Dense vectors are heavy. A 768-dim <code class="inline">float32</code> embedding is ~3 KB; a million of them is ~3 GB of vector data that has to be resident to serve low-latency kNN. Scale to tens of millions and the vector working set — not the text, not the postings — becomes the thing that decides how much RAM you rent.</p>

  <p>The standard fix is <strong>scalar quantization</strong>: store each dimension in one byte instead of four. The catch everyone worries about is recall — does compressing the vectors quietly wreck ranking quality? That is the question this recipe answers, and the answer is no.</p>

  <h2>Why XERJ</h2>
  <p>XERJ lets you opt a <code class="inline">dense_vector</code> field into <strong>scalar8</strong> (int8) quantization per field. When you do, the kNN <em>serving</em> path scores against 1-byte-per-dimension codes instead of 4-byte floats, while <code class="inline">_source</code> still returns the <strong>original</strong> vectors for retrieval. It's off by default (full float32), so you choose the precision model per field, spelled exactly like Elasticsearch's <code class="inline">int8_hnsw</code> (XERJ accepts that name as a compatibility alias — SQ8 fields are served by the exact code scan, not the HNSW ANN path, so kNN on a quantized field stays exact over the codes).</p>

  <p>On a real 128-dim corpus the recall cost is negligible: <strong>recall@10 = 0.998</strong> against the exact float32 index. That number is computed by the run below, not stipulated.</p>

  <p>What that costs in bytes <em>as an encoding</em> is 128 rather than 512 per vector, and the run measures that too — but see the note at the top: XERJ does not hold those codes resident today, so treat it as the size of the encoding, not as a saving you get.</p>

  <h2>The solution</h2>
  <p>Opt a field in at mapping time with <code class="inline">index_options.type: int8_hnsw</code>:</p>
<pre class="code">curl -sX PUT "$XERJ_URL/docs" -H 'content-type: application/json' -d '{
  "mappings": {
    "properties": {
      "title": { "type": "text" },
      "v": {
        "type": "dense_vector",
        "dims": 128,
        "similarity": "cosine",
        "index_options": { "type": "int8_hnsw" }
      }
    }
  }
}'</pre>

  <p>Index and query exactly as you would a full-precision field — nothing else changes:</p>
<pre class="code">curl -sX POST "$XERJ_URL/docs/_search" -H 'content-type: application/json' -d '{
  "knn": { "field": "v", "query_vector": [0.12, 0.08, -0.31, "..."], "k": 10 }
}'</pre>

  <p>The scores come back slightly different from an exact float32 index (that's the quantization at work — a query that exactly matches a stored vector scores ~0.99999 instead of 1.0), but the <strong>ranking is the same</strong>.</p>

  <h2>Try it</h2>
  <p><a href="https://github.com/xerj-org/xerj/blob/main/docs/examples/vector-quantization/quant_demo.py"><code class="inline">docs/examples/vector-quantization/quant_demo.py</code></a> (the mirrored <code class="inline">recipes/vector_quantization.py</code> runs the same demo) embeds the 40 real KB articles into 128-dim vectors, indexes the same vectors into a float32 index and a scalar8 index, and prints the side-by-side top hits, the measured recall@10, and the measured byte footprint of each encoding:</p>
<pre class="code">$ python3 docs/examples/vector-quantization/quant_demo.py
embedded 40 real KB articles into 128-dim vectors

indexed into `vq-none` (float32) and `vq-scalar8` (int8_hnsw / scalar8)

query: 'how do I stop an agent's context window from overflowing?'

── float32 (exact)
    0.67958  Long-context windows do not replace memory
    0.60029  p95 latency budgets for interactive RAG agents
    0.59712  SOC 2 controls that apply to vector workloads

── scalar8 (quantized)
    0.67938  Long-context windows do not replace memory
    0.60021  p95 latency budgets for interactive RAG agents
    0.59731  SOC 2 controls that apply to vector workloads

recall@10 (scalar8 vs float32 ground truth): 0.998
encoding size over 40 vecs: float32 = 20480 B (512 B/vec)  →  scalar8 = 5120 B (128 B/vec)  (4.00x smaller)

OK — recall preserved through 1-byte-per-dim codes. `_source` still holds
the originals. scalar8 changes precision, not resident memory (issue #392).</pre>

  <p>The encoding-size line is a real measurement: the run encodes every corpus vector as float32 bytes (<code class="inline">struct</code>) and as int8 codes and compares the actual byte totals — 20480 B vs 5120 B, exactly 4.00×. It is the cost of the two <em>encodings</em>, measured in the client. It is <strong>not</strong> a measurement of XERJ's resident footprint, and XERJ does not currently realise it as one (<a href="https://github.com/xerj-org/xerj/issues/392">#392</a>).</p>

  <h2>Reproduce it yourself</h2>
<pre class="code"># 1. Start XERJ (dev mode, default ES-compat port 9200)
xerj --insecure --data-dir ./data &amp;

# 2. Run the demo (stdlib-only Python 3, no packages, no API keys)
python3 docs/examples/vector-quantization/quant_demo.py</pre>
  <p><code class="inline">XERJ_URL</code> overrides the server (default <code class="inline">http://localhost:9200</code>); <code class="inline">XERJ_KB</code> overrides the KB path (default: auto-discovered <code class="inline">demo/data/ai_kb.ndjson</code>). The embedder and corpus are deterministic, so a customer should see exactly:</p>
  <ul>
    <li><code class="inline">recall@10 (scalar8 vs float32 ground truth): 0.998</code></li>
    <li><code class="inline">encoding size over 40 vecs: float32 = 20480 B (512 B/vec)  →  scalar8 = 5120 B (128 B/vec)  (4.00x smaller)</code></li>
  </ul>
  <p>These numbers are stable run-to-run (verified across repeated runs — no variance); the printed kNN scores are likewise identical each run.</p>

  <h2>Notes and limits</h2>
  <ul>
    <li><strong>Opt-in per field.</strong> Fields without <code class="inline">int8_hnsw</code> keep exact float32 scoring, byte-for-byte unchanged.</li>
    <li><strong><code class="inline">_source</code> is never quantized.</strong> Retrieval returns the vectors you indexed; only the scoring path uses the compact codes.</li>
    <li><strong><code class="inline">scalar8</code> is wired; <code class="inline">binary</code> is not yet.</strong> Binary (1-bit) quantization is rejected at startup rather than silently storing full precision.</li>
    <li><strong>Cosine is normalised</strong> before quantizing for the tightest code range; <code class="inline">dot_product</code> and <code class="inline">l2_norm</code> similarities are supported too.</li>
    <li><strong>No memory saving yet.</strong> The serving path reads the full-precision vector from <code class="inline">_source</code> and quantizes it per query, so <code class="inline">scalar8</code> costs the same RAM as float32 and a little more CPU. Tracked in <a href="https://github.com/xerj-org/xerj/issues/392">#392</a>.</li>
    <li><strong><code class="inline">scalar8</code> disables ANN.</strong> A quantized field is excluded from HNSW-served kNN and always takes the exact brute-force scan, so it is slower, not faster, than leaving the field full-precision. Also <a href="https://github.com/xerj-org/xerj/issues/392">#392</a>.</li>
    <li><strong>A <code class="inline">scalar8</code> <code class="inline">_score</code> — and the order it produces — depends on the candidate set.</strong> The codebook is fitted per query over the candidates being scored, which is what keeps an updated document from being scored on a stale codebook (<a href="https://github.com/xerj-org/xerj/issues/371">#371</a>). Three things follow, all of them measured, none of them true of a full-precision field:
      <ul>
        <li><strong>A <code class="inline">filter</code> reorders the documents it does not remove.</strong> On a 60-document 4-dim cosine corpus (30 tightly clustered <code class="inline">grp:a</code> documents, 30 spread <code class="inline">grp:b</code> ones), adding <code class="inline">filter: {"term": {"grp": "a"}}</code> — which removes only <code class="inline">grp:b</code> — returned the same 30 documents with a maximum <code class="inline">_score</code> difference of <strong>1.976e-05</strong>, but in a <strong>different order at 19 of the 30 positions</strong>. The score delta is tiny; the reordering is what a caller actually sees.</li>
        <li><strong>The trigger is the candidate set, not the <code class="inline">filter</code> keyword.</strong> Indexing one more unrelated document moves existing documents too. Same corpus, one document added far from the query: maximum <code class="inline">_score</code> difference <strong>7.100e-06</strong> over the 30 pre-existing documents, and a reordered top 10.</li>
        <li><strong>Unfiltered corpora over 1000 documents also score differently from v1.0.0-rc.17</strong>, which fitted the codebook from the first ≤1000 candidates and then cached it for the life of the process. On 1500 documents, <strong>all 40 of the top-40 <code class="inline">_score</code> values changed</strong> (maximum difference <strong>4.880e-05</strong>) and two adjacent ranks swapped. Below 1000 documents, unfiltered and never updated, scores are byte-identical to rc.17.</li>
      </ul>
      Every individual difference is bounded by SQ8's own quantization step (1/255 of the fitted per-dimension range), i.e. inside the approximation error <code class="inline">scalar8</code> already carries — but documents whose true scores are close will swap places, so treat <code class="inline">scalar8</code> ordering as stable only for a fixed corpus and a fixed filter. Elasticsearch fits its codebook per segment at index time and does not have this property; <a href="https://github.com/xerj-org/xerj/issues/392">#392</a> is what would bring XERJ in line.</li>
  </ul>
</main>
</div>

<!-- seo:footer-begin — generated by scripts/seo/fix_links.py · do not hand-edit -->
<nav class="footer-map" aria-label="Site map">
  <div class="fm-col">
    <span class="fm-h">PRODUCT</span>
    <a href="/product">PRODUCT</a>
    <a href="/architecture/">ARCHITECTURE</a>
    <a href="/benchmarks/">BENCHMARKS</a>
    <a href="/pricing/">PRICING</a>
    <a href="/security/">SECURITY</a>
    <a href="/playground">DASHBOARDS</a>
    <a href="/demo/">REAL-DATA DEMO</a>
    <a href="/aise-demo">PRODUCT TOUR</a>
  </div>
  <div class="fm-col">
    <span class="fm-h">LEARN</span>
    <a href="/answers/">ANSWERS</a>
    <a href="/compare/">COMPARE</a>
    <a href="/use-cases">USE CASES</a>
    <a href="/case-studies">CASE STUDIES</a>
    <a href="/industries/">INDUSTRIES</a>
    <a href="/blog/">BLOG</a>
  </div>
  <div class="fm-col">
    <span class="fm-h">SOLUTIONS</span>
    <a href="/solutions/">SOLUTIONS</a>
    <a href="/industries/finserv">FINANCIAL SERVICES</a>
    <a href="/industries/public-sector">PUBLIC SECTOR</a>
    <a href="/industries/healthcare">HEALTHCARE</a>
    <a href="/industries/retail">RETAIL</a>
    <a href="/use-cases/code-security-audit">AI CODE-SECURITY AUDIT</a>
    <a href="/use-cases/semantic-analytics">SEMANTIC ANALYTICS</a>
    <a href="/resources/">SALES KIT</a>
  </div>
  <div class="fm-col">
    <span class="fm-h">DOCS</span>
    <a href="/docs/">DOCUMENTATION</a>
    <a href="/docs/quickstart">QUICKSTART</a>
    <a href="/docs/install">INSTALL</a>
    <a href="/docs/migration-from-es">MIGRATE FROM ELASTICSEARCH</a>
    <a href="/docs/recipes/">RECIPES</a>
    <a href="/docs/agents/">DOCS FOR AI</a>
    <a href="/for-agents">FOR AI AGENTS</a>
    <a href="/agent-search/">AGENTIC SEARCH DEMO</a>
    <a href="/llms.txt">LLMS.TXT</a>
    <a href="https://github.com/xerj-org/xerj" rel="external">GITHUB</a>
  </div>
</nav>
<!-- seo:footer-end -->
<footer class="footer">
  <span>XERJ.AI · DOCS · V1.0.0-RC.100</span>
  <span>READ THE SOURCE</span>
  <span><a href="/">HOME</a> · <a href="/product">PRODUCT</a> · <a href="/answers/">ANSWERS</a> · <a href="/playground">DASHBOARDS</a> &middot; <a href="/llms.txt">LLMS.TXT</a></span>
</footer>

<script type="application/json" id="docs-index">[{"title":"Overview","slug":"/docs/","kicker":"START","snip":""},{"title":"Quickstart","slug":"/docs/quickstart","kicker":"START","snip":""},{"title":"Install","slug":"/docs/install","kicker":"START","snip":""},{"title":"Migrate from ES","slug":"/docs/migration-from-es","kicker":"START","snip":""},{"title":"CLI","slug":"/docs/cli","kicker":"REFERENCE","snip":""},{"title":"Config TOML","slug":"/docs/config","kicker":"REFERENCE","snip":""},{"title":"Environment","slug":"/docs/env","kicker":"REFERENCE","snip":""},{"title":"Native REST API","slug":"/docs/api-native","kicker":"REFERENCE","snip":""},{"title":"ES-compat API","slug":"/docs/api-es-compat","kicker":"REFERENCE","snip":""},{"title":"Query types","slug":"/docs/queries","kicker":"DATA MODEL","snip":""},{"title":"Aggregations","slug":"/docs/aggregations","kicker":"DATA MODEL","snip":""},{"title":"Analyzers","slug":"/docs/analyzers","kicker":"DATA MODEL","snip":""},{"title":"Vectors & kNN","slug":"/docs/vectors","kicker":"DATA MODEL","snip":""},{"title":"Ingest pipelines","slug":"/docs/ingest","kicker":"ENGINE","snip":""},{"title":"Storage & WAL","slug":"/docs/storage","kicker":"ENGINE","snip":""},{"title":"Compression","slug":"/docs/compression","kicker":"ENGINE","snip":""},{"title":"Clustering","slug":"/docs/clustering","kicker":"ENGINE","snip":""},{"title":"Metrics","slug":"/docs/metrics","kicker":"ENGINE","snip":""},{"title":"Auth & TLS","slug":"/docs/security","kicker":"ENGINE","snip":""},{"title":"Running in production","slug":"/docs/operations","kicker":"OPERATIONS","snip":""},{"title":"Backup & restore","slug":"/docs/backup-restore","kicker":"OPERATIONS","snip":""},{"title":"Upgrades","slug":"/docs/upgrades","kicker":"OPERATIONS","snip":""},{"title":"Troubleshooting","slug":"/docs/troubleshooting","kicker":"OPERATIONS","snip":""},{"title":"SIEM","slug":"/docs/playbooks/siem","kicker":"PLAYBOOKS","snip":""},{"title":"Log analytics","slug":"/docs/playbooks/log-analytics","kicker":"PLAYBOOKS","snip":""},{"title":"Vector · RAG","slug":"/docs/playbooks/vector-search","kicker":"PLAYBOOKS","snip":""},{"title":"Full-text search","slug":"/docs/playbooks/full-text","kicker":"PLAYBOOKS","snip":""},{"title":"Observability","slug":"/docs/playbooks/observability","kicker":"PLAYBOOKS","snip":""},{"title":"CLI · --config","slug":"/docs/cli#config-flag","kicker":"CLI","snip":"-c, --config <PATH>  Path to the TOML config file."},{"title":"CLI · --data-dir","slug":"/docs/cli#data-dir-flag","kicker":"CLI","snip":"-d, --data-dir <PATH>  Override [server] data_dir at the command line."},{"title":"CLI · --insecure","slug":"/docs/cli#insecure-flag","kicker":"CLI","snip":"-k, --insecure  Disable TLS and API-key auth. Dev only."},{"title":"[server] rest_port","slug":"/docs/config#server","kicker":"CONFIG","snip":"u16 · default 8080 · Native REST API listener port."},{"title":"[server] es_compat_port","slug":"/docs/config#server","kicker":"CONFIG","snip":"u16 · default 9200 · Elasticsearch-wire-format port. 0 = OS-assigned port."},{"title":"[storage] wal_sync","slug":"/docs/config#storage","kicker":"CONFIG","snip":"enum · default batched · sync · batched · async. Durability knob."},{"title":"[vector] hnsw_ef_search","slug":"/docs/config#vector","kicker":"CONFIG","snip":"u32 · default 100 · Accepted but not wired — the beam width comes from the request num_candidates."},{"title":"[vector] default_quantization","slug":"/docs/config#vector","kicker":"CONFIG","snip":"enum · none (default) · scalar8. binary is refused at startup; scalar4 is not accepted."},{"title":"GET /v1/health","slug":"/docs/api-native#health","kicker":"NATIVE API","snip":"Liveness + readiness. {status, uptime_s, version}."},{"title":"GET /v1/metrics","slug":"/docs/api-native#metrics","kicker":"NATIVE API","snip":"Prometheus text format. Scrape-friendly."},{"title":"POST /v1/indices/:name/turbo-ingest","slug":"/docs/api-native#turbo-ingest","kicker":"NATIVE API","snip":"NDJSON bulk path with parallel tokenization."},{"title":"POST /v1/indices/:name/search","slug":"/docs/api-native#search","kicker":"NATIVE API","snip":"Unified query endpoint — full-text, term, range, KNN, hybrid."},{"title":"POST /:index/_bulk","slug":"/docs/api-es-compat#bulk","kicker":"ES API","snip":"NDJSON bulk. Actions: index, create, update, delete."},{"title":"POST /:index/_search","slug":"/docs/api-es-compat#search","kicker":"ES API","snip":"ES query DSL. Drop-in migration path."},{"title":"match_all","slug":"/docs/queries#match-all","kicker":"QUERY","snip":"Everything in the index."},{"title":"bool","slug":"/docs/queries#bool","kicker":"QUERY","snip":"must / should / must_not / filter clauses."},{"title":"knn","slug":"/docs/queries#knn","kicker":"QUERY","snip":"Vector kNN — HNSW-served when unfiltered (exact-rescored); filtered runs the exact scan."},{"title":"semantic","slug":"/docs/queries#semantic","kicker":"QUERY","snip":"Embed the query text at query time and search the vector field."},{"title":"hybrid","slug":"/docs/queries#hybrid","kicker":"QUERY","snip":"RRF / linear fusion of sub-queries."},{"title":"terms aggregation","slug":"/docs/aggregations#terms","kicker":"AGG","snip":"Exact, pre-computed terms histograms over keyword fields."},{"title":"date_histogram","slug":"/docs/aggregations#date-histogram","kicker":"AGG","snip":"Time-bucketed counts."},{"title":"kNN serving","slug":"/docs/vectors#knn-query","kicker":"VECTORS","snip":"Unfiltered kNN via persisted HNSW + exact rescore; filtered kNN exact scan."},{"title":"SQ8 quantization","slug":"/docs/vectors#quantization","kicker":"VECTORS","snip":"8-bit scalar quantization. int8 scoring precision, not a memory saving (#392)."},{"title":"turbo-ingest","slug":"/docs/ingest#turbo","kicker":"INGEST","snip":"NDJSON bulk path. Parallel tokenization across cores."},{"title":"delta-of-delta","slug":"/docs/compression#delta-of-delta","kicker":"COMPRESSION","snip":"Timestamp encoding. Apache/Nginx/ISO auto-detect."},{"title":"dictionary encoding","slug":"/docs/compression#dictionary","kicker":"COMPRESSION","snip":"Up to 256 unique strings per column."},{"title":"api_key","slug":"/docs/security#api-key","kicker":"SECURITY","snip":"Static admin key. Auto-generated on first run."},{"title":"SIEM playbook","slug":"/docs/playbooks/siem","kicker":"PLAYBOOK","snip":"Terms agg over source IPs for SIEM, with detection rules."},{"title":"RAG playbook","slug":"/docs/playbooks/vector-search","kicker":"PLAYBOOK","snip":"Embeddings, kNN, hybrid search, BM25 + kNN fusion."},{"title":"[cluster] enabled","slug":"/docs/config#cluster","kicker":"CONFIG","snip":"bool · default false · Enable multi-node Raft cluster mode."},{"title":"[cluster] peers","slug":"/docs/config#cluster","kicker":"CONFIG","snip":"array · \"host:port=host:port\" · label must equal address (#1171)."},{"title":"[cluster] port","slug":"/docs/config#cluster","kicker":"CONFIG","snip":"u16 · default 9300 · Intra-cluster Raft + search port."},{"title":"[embedding] default_endpoint","slug":"/docs/config#embedding","kicker":"CONFIG","snip":"string · OpenAI-compatible embeddings URL for auto-embedding on ingest."},{"title":"[embedding] default_model","slug":"/docs/config#embedding","kicker":"CONFIG","snip":"string · Model name (e.g. text-embedding-3-small, nomic-embed-text)."},{"title":"Raft","slug":"/docs/clustering","kicker":"CLUSTER","snip":"Embedded Raft — leader election, log replication, metadata only."},{"title":"Node draining","slug":"/docs/clustering","kicker":"CLUSTER","snip":"POST /v1/cluster/nodes/:id/drain before maintenance."},{"title":"Region split/merge","slug":"/docs/clustering","kicker":"CLUSTER","snip":"Range partitions auto-split at byte threshold, merge back when small."},{"title":"Systemd unit","slug":"/docs/operations","kicker":"OPS","snip":"Production unit file with hardening directives."},{"title":"Readiness probe","slug":"/docs/operations","kicker":"OPS","snip":"GET /v1/health/ready — only 200 once WAL replay is complete."},{"title":"Config reload","slug":"/docs/operations","kicker":"OPS","snip":"SIGHUP reloads most keys without restart."},{"title":"Hot backup","slug":"/docs/backup-restore","kicker":"OPS","snip":"rsync the data dir after POST /v1/admin/flush."},{"title":"S3 backup","slug":"/docs/backup-restore","kicker":"OPS","snip":"POST /v1/admin/backup → any S3-compatible endpoint."},{"title":"Rolling upgrade","slug":"/docs/upgrades","kicker":"OPS","snip":"Drain → stop → upgrade → start → activate, per node."},{"title":"Merge pressure","slug":"/docs/troubleshooting","kicker":"OPS","snip":"Segment count climbing → raise merge.max_concurrent + io_rate."},{"title":"Cluster flapping","slug":"/docs/troubleshooting","kicker":"OPS","snip":"Leader changes every few seconds → raise cluster.tick_ms."},{"title":"WAL warnings after a crash","slug":"/docs/troubleshooting","kicker":"OPS","snip":"Repaired on startup; no manual repair command."},{"title": "Recipes", "slug": "/docs/recipes/", "kicker": "RECIPES", "snip": "Task-oriented AI & search recipes, each verified end-to-end against a live XERJ."},{"title": "Semantic search & RAG", "slug": "/docs/recipes/semantic-search-rag", "kicker": "RECIPE", "snip": "Retrieval by meaning with semantic_text — auto-embed on ingest, no separate vector DB."},{"title": "Agent memory", "slug": "/docs/recipes/agentic-memory", "kicker": "RECIPE", "snip": "Long-term memory for agents via /_memory — store, recall, forget, per-agent isolation."},{"title": "Passage retrieval", "slug": "/docs/recipes/passage-retrieval", "kicker": "RECIPE", "snip": "Auto-chunk long docs; best-passage (max-sim) scoring — 98% top-3."},{"title": "Vector search · kNN", "slug": "/docs/recipes/vector-search-knn", "kicker": "RECIPE", "snip": "Nearest-neighbor over dense_vector — HNSW-served unfiltered, exact with filters."},{"title": "Vector quantization", "slug": "/docs/recipes/vector-quantization", "kicker": "RECIPE", "snip": "scalar8 (accepts ES int8_hnsw; exact code scan) — int8 scoring at recall@10 = 0.998. Precision, not memory."},{"title": "Hybrid search", "slug": "/docs/recipes/hybrid-search", "kicker": "RECIPE", "snip": "Keyword + vector in one query — RRF fusion."},{"title": "Log analytics", "slug": "/docs/recipes/log-analytics", "kicker": "RECIPE", "snip": "Raw logs to dashboards — error rates, p95 latency, top services."},{"title": "Anomaly detection", "slug": "/docs/recipes/anomaly-detection", "kicker": "RECIPE", "snip": "Statistical _ml detectors that flag spikes in metrics/logs."},{"title": "Anomaly datafeeds", "slug": "/docs/recipes/continuous-anomaly-datafeeds", "kicker": "RECIPE", "snip": "Live _ml datafeed re-scores an index on a timer, stores anomaly records."},{"title": "Migrate from ES", "slug": "/docs/recipes/migrate-from-elasticsearch", "kicker": "RECIPE", "snip": "Point your existing ES client at XERJ — same wire, change the URL."}]</script>

<script>
  // IIFE — js-auth navigates between decrypted pages in the same
  // JS realm, so const bindings must be lexically scoped.
  (function () {
    const root = document.documentElement;

    // Client-side docs search — filters the inline JSON index.
    const searchEl    = document.getElementById('docsSearch');
    const inputEl     = document.getElementById('docsSearchInput');
    const resultsEl   = document.getElementById('docsSearchResults');
    const indexJson   = document.getElementById('docs-index');
    if (searchEl && inputEl && resultsEl && indexJson) {
      let index = [];
      try { index = JSON.parse(indexJson.textContent || '[]'); } catch {}
      const render = (q) => {
        const query = q.trim().toLowerCase();
        if (!query) { searchEl.classList.remove('active'); resultsEl.innerHTML = ''; return; }
        const scored = [];
        for (const e of index) {
          const hay = (e.title + ' ' + (e.kicker || '') + ' ' + (e.snip || '')).toLowerCase();
          if (hay.includes(query)) {
            let score = 0;
            if (e.title.toLowerCase().includes(query)) score += 10;
            if ((e.kicker || '').toLowerCase() === query.toUpperCase()) score += 5;
            if ((e.snip || '').toLowerCase().includes(query)) score += 1;
            scored.push({ e, score });
          }
        }
        scored.sort((a, b) => b.score - a.score);
        const top = scored.slice(0, 20);
        if (top.length === 0) {
          resultsEl.innerHTML = '<div class="no-results">NO MATCHES — TRY A SHORTER QUERY</div>';
        } else {
          resultsEl.innerHTML = top.map(({e}) => {
            const kicker = e.kicker || '';
            const snip = e.snip ? '<div class="snip">' + escHtml(e.snip) + '</div>' : '';
            return '<a href="' + e.slug + '">' +
              '<span class="kicker">' + escHtml(kicker) + '</span>' +
              '<span class="title">' + escHtml(e.title) + '</span>' +
              snip + '</a>';
          }).join('');
        }
        searchEl.classList.add('active');
      };
      const escHtml = (s) => String(s).replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;').replace(/"/g, '&quot;');
      inputEl.addEventListener('input', (e) => render(e.target.value));
      inputEl.addEventListener('focus', () => { if (inputEl.value) render(inputEl.value); });
      document.addEventListener('keydown', (e) => {
        if ((e.key === 'k' || e.key === 'K') && (e.metaKey || e.ctrlKey)) {
          e.preventDefault();
          inputEl.focus();
          inputEl.select();
        }
        if (e.key === 'Escape' && document.activeElement === inputEl) {
          inputEl.value = '';
          searchEl.classList.remove('active');
          resultsEl.innerHTML = '';
          inputEl.blur();
        }
      });
    }
  })();
</script>

</body>
</html>