--- title: "Search a help-center you saved as HTML" canonical: "https://xerj.org/answers/search-html-export" updated: "2026-08-21" source: "content/answers/search-html-export.md" --- # I saved a help-center as HTML. How do I search it like the real help-center search? **TL;DR** — XERJ `autoindex` indexes a static HTML export straight from local disk. In our 3-page capture XERJ extracted a `title` and a `headings` array from each page, and one `match` query on `body` returned 3 hits. A network watch observed 0 non-loopback peers during the run. ## Agent prompt ```text Act as a coding agent. Read https://xerj.org/llms.txt, start a XERJ node with --insecure, run xerj autoindex ./html-export --prefix web on a folder of saved HTML pages, then POST a match query on body to /web-*/_search and report ax_path, title and the headings array for every hit without fetching any URL. ``` ## Commands ### Command 1 Note: Index a folder of saved HTML pages from local disk. ```sh xerj autoindex ./html-export --url http://127.0.0.1:9410 --prefix web --state-dir ./state-web --progress plain --disable-feedback ``` ### Command 2 Note: Search the page text and return the title and headings of each hit. ```sh curl -s -XPOST 'http://127.0.0.1:9410/web-*/_search' -H 'content-type: application/json' -d '{"query":{"match":{"body":"checkpoint journal"}},"size":10,"_source":["ax_path","title","headings","section"],"track_total_hits":true}' ``` ## Index the export from disk `xerj autoindex` indexes the HTML files that are already on disk and fetches no URL. XERJ has no fetcher and follows no `href`, so a saved site export is the input, not a live site. The command below indexes a folder of 3 static pages. ```sh xerj autoindex ./html-export --url http://127.0.0.1:9410 --prefix web --state-dir ./state-web --progress plain --disable-feedback ``` ## Fields the HTML family produced The HTML family produced `title`, `headings` and `body`, on top of the 7 `ax_*` provenance fields. The `headings` field is an array of the heading text in document order, so `recovery.html` carried `["Recovery procedure", "Overview"]`. That field set distinguishes HTML from plain text in XERJ. A Markdown file lands in the `txt-prose` family and gets no `headings` array, as [the Markdown answer](/answers/index-markdown-into-elasticsearch-api) shows. ## The query and its 3 hits One `match` query on `body` for `checkpoint journal` returned 3 hits, one per page, ranked by BM25. The captured `_source` carried `ax_path`, `title` and `headings` for every hit. ```sh curl -s -XPOST 'http://127.0.0.1:9410/web-*/_search' -H 'content-type: application/json' -d '{"query":{"match":{"body":"checkpoint journal"}},"size":10,"_source":["ax_path","title","headings","section"],"track_total_hits":true}' ``` | `ax_path` | `title` | `headings` | `_score` | | --- | --- | --- | --- | | `recovery.html` | Recovery procedure | `["Recovery procedure", "Overview"]` | `0.61400104` | | `index.html` | Export index | `["Runbook export", "Overview"]` | `0.57417387` | | `glossary.html` | Glossary | `["Glossary", "Overview"]` | `0.14597225` | ## Proof that no URL was fetched A network watch over the XERJ node and the whole harness process tree observed 0 non-loopback peers for the whole run. The watch polled `/proc/net/tcp` and `/proc/net/tcp6` and cross-referenced the socket inodes of up to 4 process ids in the watched tree. The method has one honest limit, and the capture states it. The watch is a sampler, not a packet capture, and it took 4 samples at a 0.05 second interval across 0.32 seconds. A connection that opens and closes inside one gap can escape the sampler. ## Limits worth knowing before you start XERJ indexes files from a filesystem walk on 1 single-node process, so the export must exist on disk first. A hidden dotfile and a dotted directory are always skipped, and `--no-ignore` does not turn that off. XERJ elects `body` as a `semantic_text` field on document datasets. The default embedder is lexical feature hashing, and the neural embedder is opt-in through `--embed-mode neural`. A `match` query on a `semantic_text` field runs BM25, not kNN. ## FAQ ### I saved a bunch of docs pages as HTML. How do I search that? Index the folder from disk and send a `match` query on `body`. The HTML family produces `title`, `headings` and `body`, so a hit names the page and its section headings. ### How do I search my Confluence HTML export? The same way. XERJ reads the saved files from disk and fetches no URL, and the Confluence export article covers that space layout in detail. ### How do I search a wget of vendor docs? Point `autoindex` at the download directory. Every page becomes a document with `title`, `headings` and `body`, and one `match` query on `body` covers the whole mirror. ### Does site navigation drown the real hits? It can. Every saved page repeats its nav and footer text in `body`, so a common word matches many pages. Query a distinctive phrase, or read `title` and `headings` to keep the answer inside the article text. ### Does XERJ download the pages itself? No. XERJ indexes HTML files that are already on disk and fetches no URL. Our network watch observed 0 non-loopback peers during the run. ### Can I see which headings a page has? Yes. Ask for `headings` in `_source`. In our capture `recovery.html` returned `["Recovery procedure", "Overview"]` in document order. ### Is the zero-network result a packet capture? No. The capture is a sampler at a 0.05 second interval, so a connection opening and closing inside one gap would be missed. The capture states that limit. ## Related - [I have a folder of PDFs, Word docs, and markdown. How do I search all of them at once?](/answers/search-file-contents-in-a-folder) - [How do I search a Confluence export?](/answers/search-confluence-html-export) - [Can ChatGPT search a folder on my laptop, or do I need something else?](/answers/give-chatgpt-claude-local-file-access) - [Pagefind vs indexing a local HTML dump for an agent?](/compare/xerj-vs-pagefind)