These pages show where XERJ is used, each with measured results and
limits. Start with on-prem RAG with Cohere, a WordPress security audit,
a self-audit of the XERJ engine, log analytics, or an Elasticsearch
replacement. Audit queries belong on the
AI security review
use case and the two case studies below. This page is the index.
Capability charts against Elasticsearch 8.13 are reproducible from
demo/playbooks/SCORECARD.md.
Find your XERJ workflow
Pick an answer above.
On-prem AI with Cohere
Regulations, manuals and a live 200,000-row Postgres table on one CPU server, with citations.
Read the use caseZero-token AI search
Point XERJ at a folder of code, PDFs or Markdown and get the exact files back.
Read the use caseSecond brain
xerj brain turns a folder of notes into an evidence-carrying knowledge graph.
Read the use caseAI security review
Index a codebase as security facts, then read only the violators.
Read the use caseSecurity analytics
SOC filters, term lookups and aggregations on a single XERJ node.
Read the use caseOperational intelligence
One ingest endpoint, one storage engine and one retention number for logs.
Read the use caseUnified observability
Logs, metrics and traces in one engine with one query language.
Read the use caseElasticsearch replacement
Same :9200 API, query DSL and client libraries on one binary.
Read the use caseTwo reproducible whitebox security audits with XERJ as the agent's retrieval substrate, one on a large third-party codebase, one on XERJ's own engine. Both real, both measured, each its own full case study.
Regulations, claims manuals and a live 200,000-row Postgres claims table on one CPU server. Documents searchable in 7.9 s, database changes searchable in about 20 ms, plain-language questions answered with evidence, Cohere Rerank and Command on top with citations back to file and page. A worked example on public data; the Cohere calls ran against a mock until a key is available.
An AI agent security-audited real WordPress core, 1,492 PHP files, ~619k lines, using XERJ as its retrieval substrate: index once as security facts (11,990 functions, 1,343 hooks), query the bug shape, read only the survivors. It swept 226 handlers for missing-capability IDOR (0 found, core is hardened) and surfaced the real wp_http_validate_url SSRF gap (169.254 cloud-metadata) reachable unauthenticated. The point is the method and its economics: the whole audit cost ~26K tokens against ~5.2M just to load the tree.
TOKEN COST TO THE ANSWER · NOT LATENCY · CORE IS HARDENED. 0 IDOR ACROSS 226 HANDLERS; THE ONE REAL WEAKNESS (SSRF) IS STATED PLAINLY · SOURCES IN docs/case-studies/wordpress-security-audit →
The stronger proof: XERJ turned the method on its own Rust engine (197 files) and found and fixed real vulnerabilities. A call-graph query grep cannot express surfaced an unauthenticated /_sql stack-overflow that aborts the process; a five-stage instruction/data pass added a search-template injection that leaks private documents and a snapshot path-traversal. The three application-layer highs are fixed on the branch, each proven by crashing, then not-crashing, a live server. Against six known-fixed bugs the substrate recalled 6/6 where grep managed 4/6, at 171× fewer tokens.
TOKEN COST TO THE ANSWER · NOT LATENCY · RECALL MEASURED AGAINST 6 KNOWN-FIXED BUGS (grep CAN'T FIND MISSING CODE); 6/6 IS IN-SAMPLE. SEE THE CAVEAT · METHODOLOGY + ALL 8 FINDINGS IN docs/case-studies/xerj-self-audit →
Reproducible capability benchmarks from demo/playbooks/SCORECARD.md: measured numbers on a real query matrix, not user stories.
Your SOC team spends more time waiting for dashboards than investigating threats. The battery on the right is the reproducible XERJ-vs-Elasticsearch query matrix, bool filters, range scans, term lookups and aggregations, measured on a 1M-doc corpus. XERJ is the yellow bar; Elasticsearch 8.13 is the gray one.
P50 · 1M-DOC LLM-TELEMETRY CORPUS · FULL 91-CELL MATRIX AT xerj.org/benchmarks →
Ship logs at line rate. Query hot data in milliseconds. No ILM policy puzzles, no shard tuning, no Logstash pipeline. The chart shows a 24-hour diurnal ingest pattern. XERJ sustains it with zero write-stalls, and ingests 1.5-1.7× faster than Elasticsearch in the reproducible benchmark.
Hybrid semantic + keyword search in one query. The cluster plot shows 830K real queries grouped into six intent clusters. RAG retrieval dominates. Every cluster is one index query away, not a separate Pinecone call. The benchmark below is the reproducible XERJ-vs-Elasticsearch latency for kNN and text retrieval.
P50 · 1M-DOC LLM-TELEMETRY CORPUS · FULL 91-CELL MATRIX AT xerj.org/benchmarks →
Same API on port 9200. Same query DSL. Same client libraries. The comparison on the right is the reproducible query-latency head-to-head. Yellow is XERJ; gray is Elasticsearch 8.13 on the identical workload.
P50 · 1M-DOC LLM-TELEMETRY CORPUS · FULL 91-CELL MATRIX AT xerj.org/benchmarks →
Logs, traces, and metrics in one store, one binary instead of a Grafana-Loki-Tempo-Mimir stack to babysit. OTLP in, Prometheus out. The chart on the right is the reproducible query-latency head-to-head vs Elasticsearch 8.13 on log-shaped data.
P50 · 1M-DOC LLM-TELEMETRY CORPUS · FULL 91-CELL MATRIX AT xerj.org/benchmarks →
Leave a work email and we'll reply within one business day with a private binary, the reproduction scripts, and a 45-minute walkthrough.
Install XERJ (docs: https://xerj.org/llms.txt) and pick a use case from https://xerj.org/use-cases that matches my project. Then follow that page's measured setup.
This index collects published XERJ use cases: on-prem RAG with Cohere, source-code audits, log analytics, observability, AI search, and Elasticsearch replacement. Each card links to a page with measured results and limits.
The AI security review use case is at /use-cases/code-security-audit. The WordPress audit and the XERJ self-audit are case studies linked from this page.
An agent audited WordPress core (1,492 PHP files) with XERJ as retrieval. Token cost to the answer is on that case study, including a 199 times comparison against reading the tree.
XERJ indexed its own Rust engine and found real issues. Against six known-fixed bugs, recall was 6/6, against 4/6 for grep, with 171 times fewer tokens. Three application-layer highs were fixed on the branch.