XERJ Search & Agent Memory Expert / Panduan Ahli Mesin Pencari & Memori Agen XERJ

English | Bahasa Indonesia

---

<a name="english"></a>

English

Orchestration & Integration

• Upstream Orchestrators: Works under brainstorming and zero-to-prod-orchestrator (Phase 3 - Database & Search Architecture and Phase 4 - Backend APIs & AI Agents) as the unified retrieval and memory engine.
• Downstream Cooperating Skills:
  • search-engine-expert — Integrating Elasticsearch wire-compatible endpoints, query DSLs, BM25 ranking, and faceted search.
  • vector-db-rag-expert — Hybrid search (BM25 + vector RRF), dense vector indexing, and HNSW graph retrieval.
  • agentic-memory-architect — Persistent agent episodic memory with /_memory/{ns} and namespace partitioning.
  • mcp-server-architect — Configuring and routing the 13 built-in Model Context Protocol (MCP) tools via xerj mcp.
  • anti-slop — Enforcing honest claims, verified benchmark numbers, and zero-hallucination citations.
  • session-memory-manager — Store XERJ index definitions and agent memory configuration checkpoints.

Description

Expert guide for installing, running, autoindexing, and querying XERJ — the unified, AI-native search engine, vector database, agent memory store, and log analytics engine written from scratch in Rust (Apache-2.0). XERJ replaces the complex retrieval stack an AI system normally assembles (Elasticsearch + vector database + embedding service + memory store) with a single native binary, speaking both native REST and Elasticsearch 8.x wire protocol (0-failure conformance gate) and exposing 13 native MCP tools.

Key Capabilities

┌────────────────────────────────────────────────────────────────────────┐
│                              XERJ ENGINE                               │
│                                                                        │
│  ┌───────────────────────┐  ┌───────────────────────┐  ┌────────────┐  │
│  │ xerj autoindex <dir>  │  │ xerj autoindex map    │  │ /_memory/  │  │
│  │ Zero-config indexing  │  │ Self-describing data  │  │ Long-term  │  │
│  │ Sniffs & types data   │  │ Ready-to-use queries  │  │ Agent Mem  │  │
│  └───────────────────────┘  └───────────────────────┘  └────────────┘  │
│                                                                        │
│  ┌───────────────────────┐  ┌───────────────────────┐  ┌────────────┐  │
│  │ xerj mcp (13 Tools)   │  │ ES 8.x Wire Protocol  │  │ xerj code  │  │
│  │ Native stdio proxy    │  │ :9200 compatible      │  │ Reference  │  │
│  │ Full AI Agent toolkit │  │ 0-failure YAML gate   │  │ Coding Hub │  │
│  └───────────────────────┘  └───────────────────────┘  └────────────┘  │
└────────────────────────────────────────────────────────────────────────┘

1. Zero-Config Autoindex (xerj autoindex <folder>):
   • Walks, sniffs content (not just file extensions), samples types, and builds typed, queryable indices.
   • Always run with --dry-run first to estimate size and see ignore accounting.
   • Exit codes: 0 complete, 3 completed-with-junk (success; corpus is committed), 4 needs user decision.
2. Data Map (xerj autoindex map):
   • Inspects the autoindex catalog and outputs dataset schemas, cardinality, semantic fields, and ready-to-copy curl queries.
3. Agent Episodic Memory (/_memory/{ns}):
   • Built-in persistent namespace-isolated agent memory for stateful conversational agents and long-running swarms.
4. Embedded MCP Server (xerj mcp):
   • Zero-dependency Model Context Protocol stdio server inside the same binary.

---

The 13 Native MCP Tools

Tool Name	Type	Description
xerj_search	Thin Proxy	Full-text BM25 search across indexed fields.
xerj_semantic_search	Thin Proxy	Semantic search (lexical hash by default; neural Candle BERT when --embed-mode neural).
xerj_vector_search	Thin Proxy	Dense vector search via persisted HNSW graph with exact rescoring.
xerj_hybrid_search	Thin Proxy	Reciprocal Rank Fusion (RRF) combining BM25 keyword and dense vector scores.
xerj_memory_store	Thin Proxy	Writes an agent memory observation into namespace /_memory/{ns}.
xerj_memory_recall	Thin Proxy	Recalls relevant agent memory observations from namespace /_memory/{ns}.
xerj_brain_overview	Thin Proxy	High-level topology overview of the deterministic link graph.
xerj_brain_ego	Thin Proxy	Queries the ego-network and outgoing/incoming edges for a node.
xerj_brain_link	Thin Proxy	Adds a directed semantic edge between two entities.
xerj_brain_unlink	Thin Proxy	Removes a directed edge between two entities.
xerj_plan	Thin Proxy	Invokes the deterministic rule-based query planner (POST /_ask).
xerj_map	Catalog Reader	Returns the autoindex catalog and per-index field mappings.
xerj_code_search	Rich Pipeline	Reference-coding search with staleness checks and licence attribution.

---

MCP Client Configurations

1. Antigravity IDE (.agents/mcp.json or ~/.gemini/config/mcp.json)

{
  "mcpServers": {
    "xerj": {
      "command": "xerj",
      "args": ["mcp"],
      "env": {
        "XERJ_URL": "http://localhost:9200"
      }
    }
  }
}

Note for Windows: If xerj is not in system PATH, use absolute path: %LOCALAPPDATA%\\Programs\\xerj\\xerj.exe. Add "XERJ_AUTH": "ApiKey <key>" if auth is enabled.

2. Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "xerj": {
      "command": "/usr/local/bin/xerj",
      "args": ["mcp"],
      "env": {
        "XERJ_URL": "http://127.0.0.1:9200"
      }
    }
  }
}

3. Cursor (.cursor/mcp.json)

{
  "mcpServers": {
    "xerj": {
      "command": "xerj",
      "args": ["mcp"],
      "env": {
        "XERJ_URL": "http://localhost:9200"
      }
    }
  }
}

---

Invariants & Anti-Slop Honesty Rules

1. Embedding Mode Transparency: The default embedder is lexical feature-hashing. Never describe default output as neural or semantic transformer embeddings unless the node runs with --embed-mode neural.
2. Deterministic Link Graph: xerj_brain_* tools operate a deterministic link index, not a full graph database.
3. Verified Measurements: Always quote measured benchmarks (e.g. from scripts/bench/ or demo/playbooks/SCORECARD.md); never fabricate unverified speedup ratios.

---

<a name="bahasa-indonesia"></a>

Bahasa Indonesia

Integrasi Orkestrasi

• Orkestrator Hulu: Beroperasi di bawah arahan brainstorming dan zero-to-prod-orchestrator (Fase 3 - Arsitektur Database & Search dan Fase 4 - Backend API & Agen AI) sebagai mesin retrieval dan memori terpadu.
• Skill Hilir & Kerja Sama:
  • search-engine-expert — Integrasi endpoint kompatibel Elasticsearch, query DSL, skor BM25, dan faceted search.
  • vector-db-rag-expert — Pencarian hybrid (BM25 + vector RRF), pengindeksan vektor densitas tinggi, dan graf HNSW.
  • agentic-memory-architect — Penyimpanan memori episodik jangka panjang agen melalui /_memory/{ns}.
  • mcp-server-architect — Konfigurasi dan perutean 13 alat Model Context Protocol (MCP) native via xerj mcp.
  • anti-slop — Penegakan klaim jujur, verifikasi angka benchmark, dan atribusi tanpa halusinasi.
  • session-memory-manager — Menyimpan definisi indeks XERJ dan checkpoint konfigurasi memori agen.

Deskripsi

Panduan ahli untuk instalasi, pengoperasian, pengindeksan otomatis (autoindex), dan pencarian menggunakan XERJ — mesin pencari AI-native, vector database, memori agen, dan analitik log terpadu yang dibangun dari nol dalam bahasa Rust (Apache-2.0). XERJ menggantikan seluruh tumpukan retrieval terpisah (Elasticsearch + vector database + embedding service + memory store) menjadi satu binary mandiri, mendukung protokol Elasticsearch 8.x (gerbang conformance CI 0-failure) dan menyediakan 13 tool MCP bawaan.

Kemampuan Utama

1. Autoindex Tanpa Konfigurasi (xerj autoindex <folder>):
   • Memindai konten (berdasarkan sniffing data, bukan sekadar ekstensi), mengambil sampel tipe data, dan membangun indeks siap cari secara otomatis.
   • Selalu jalankan dengan --dry-run terlebih dahulu untuk estimasi kapasitas dan melihat aturan pengabaian file (ignore rules).
   • Kode keluar: 0 selesai sempurna, 3 selesai dengan berkas lewatan (junk files, status sukses), 4 membutuhkan persetujuan pengguna.
2. Peta Data (xerj autoindex map):
   • Membaca katalog dan memaparkan skema dataset, kardinalitas kolom, kolom semantik, serta contoh query curl siap pakai.
3. Memori Jangka Panjang Agen (/_memory/{ns}):
   • Penyimpanan memori berisolasi namespace bawaan untuk agen percakapan dan multi-agent swarm.
4. Server MCP Native (xerj mcp):
   • Menyediakan 13 tool MCP langsung dari binary tunggal tanpa dependensi tambahan.

Ringkasan 13 Tool MCP XERJ

• xerj_search: Pencarian full-text BM25.
• xerj_semantic_search: Pencarian teks semantik (lexical hash default, neural BERT saat flag aktif).
• xerj_vector_search: Pencarian dense vector berbasis graf HNSW dengan exact rescoring.
• xerj_hybrid_search: Penggabungan skor BM25 dan vektor melalui RRF (Reciprocal Rank Fusion).
• xerj_memory_store & xerj_memory_recall: Penyimpanan dan pemanggilan memori agen pada /_memory/{ns}.
• xerj_brain_*: Manajemen graf tautan deterministik.
• xerj_plan: Perencana query otomatis berdasar aturan (POST /_ask).
• xerj_map: Peta data skema indeks otomatis.
• xerj_code_search: Pencarian referensi kode dengan perlindungan lisensi dan staleness check.