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.
Retrieve several IDs or questions in one call
xerj batch-search and the MCP tool xerj_batch_search share one adapter.
They retrieve evidence from a node that is already running; neither starts a
node, downloads a corpus, writes data, or generates an answer.
Save this as queries.json, replacing the index and field with those in your
data map. id must be an indexed keyword field for the identifier lookup.
There is no guessed ID field and no tokenized-text fallback for exact IDs.
{
"index": "advisories",
"size": 3,
"queries": [
{"key": "known", "identifier": "CVE-2026-12345", "field": "id"},
{"key": "question", "query": "when should we rotate keys"},
{"key": "structured", "query": {"match": {"body": "rotation"}}}
]
}
xerj batch-search --input queries.json --url http://127.0.0.1:9200
# Or pipe the same JSON:
cat queries.json | xerj batch-search --input -
For MCP, call xerj_batch_search with that JSON object as arguments.
XERJ_URL, XERJ_AUTH, explicit --url/--auth and loopback admin-key discovery
work as for xerj mcp. The node must already be running. prefix can replace
index; combining both is an error. With neither, the target is ax-*.
The response has one results entry for each input, in the same order. Its
zero-based position is authoritative; optional key values may repeat.
| Status | Meaning |
|---|---|
hits, abstain: false |
Retrieved hits, each with citation and raw hit evidence. A hit is not a generated answer or a calibrated relevance guarantee. |
abstain, abstain: true, reason: no_hits |
The item completed with an empty hit list. This is retrieval abstention, not a confidence judgment. |
error |
Invalid item, unsupported exact-field mapping, failed/partial search, malformed response, or transport failure. Never relabeled as a miss. |
Citations always retain _index and _id. Available ax_path/path,
ax_locator, ax_file, stored line/page/cell and calculated match_line fields
are included too. No file line is inferred from a JSONL record ordinal.
There is no deduplication across questions, even for identical inputs.
Limits: 1–50 items, size 1–50 per item (default 3), a 1 MiB serialized input,
256 UTF-8 bytes per optional key, a 32 MiB HTTP response, and the MCP client's
30-second HTTP timeout. Locally invalid items do not reach the node. The other
items use one _msearch request and, when needed, one shared mapping read.
An unreadable mapping fails exact-ID items; plain text retains the established
search-query fallback. Exact ID fields must be flat and keyword-mapped wherever
present, with indexing enabled. Nested fields are outside this first version.
max_tokens defaults to 8,000 when omitted or null, for both MCP and
xerj batch-search. Pass a larger integer to request more. It caps the entire
successful response JSON, using the
same UAX #29 word counter as the other MCP search tools. This is not BPE.
If binding, the adapter removes raw hit bodies starting at later items,
then extra hits; it retains every item outcome and at least one citation per
matched item. hits_omitted counts omitted hits, and _token_budget names the
counter and limit. A nonbinding budget leaves the response byte-identical.
If the remaining statuses, errors and citations cannot fit, the whole call
returns an explicit error naming the minimum required budget. Split the batch
or increase the budget. Errors are outside the budget and are never truncated. This is a word budget,
not a byte cap or a guarantee that output fits a host's display limit.
CLI exit codes: 0 all items completed (including abstentions); 3 one or more
per-item errors (the JSON still contains all outcomes); 2 invalid CLI/input
envelope; 1 a batch-wide execution or budget error. Treat exit 3 as a partial
result to inspect, not a reason to discard the successful items.
The batch path uses _msearch rather than repeated _search calls. Plain-text
items reuse the mapping-aware search query and code-path demotion; the engine's
_msearch renderer returns stored evidence, not _search's extracted _passage
field. Rerank/judge and per-item scopes/options are not exposed here. No latency,
model-token savings or confidence claim is made by this recipe.
For a pinned rust-vulns retrieval evaluation comparing 50-item batches with one-item calls over both CLI and persistent MCP, see the batch retrieval harness. It measures labelled ID lookup, known-item text hit@3, retrieval abstentions, warm task latency and payload BPE tokens; these are not generated-answer or complete agent-task cost measurements.
This is the batch retrieval part of #1801. Full answer/abstain calibration and alias-aware ID discovery remain separate work; the adapter and retrieval evaluation do not close that issue.