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.

test(vectors): lock semantic_text built-in embedder end-to-end (lexical, not neural)

xerj-org1f4522a76fc4
diff --git a/engine/tests/es-compat-yaml/yaml/vectors/50_semantic_text_builtin.yml b/engine/tests/es-compat-yaml/yaml/vectors/50_semantic_text_builtin.yml
new file mode 100644
index 000000000..4d365bf35
--- /dev/null
+++ b/engine/tests/es-compat-yaml/yaml/vectors/50_semantic_text_builtin.yml
@@ -0,0 +1,83 @@
+# semantic_text built-in embedder — end-to-end regression.
+#
+# COMPAT_MATRIX / lexical-not-neural note:
+# XERJ's `semantic_text` field auto-embeds text on ingest with a zero-config,
+# deterministic BUILT-IN embedder (xerj_ai::local::local_embed) that hashes
+# word unigrams + intra-word character trigrams into an L2-normalised vector.
+# It captures LEXICAL / vocabulary overlap, NOT deep neural semantics. This is
+# an intentional, disclosed limitation (see README.md "Vector & semantic search
+# notes" and ROADMAP.md — a bundled neural model is future work). When an
+# external OpenAI-compatible `embedding.default_endpoint` is configured the
+# engine uses that proxy at ingest and query time instead, at higher quality.
+#
+# What this test locks (the guarantee the built-in embedder makes):
+# * `semantic_text` works with ZERO external config — create → index → query.
+# * The query text is embedded with the SAME built-in embedder used at ingest,
+# so a paraphrase sharing vocabulary ranks ABOVE topically-unrelated text.
+# * `k` bounds the neighbour pool (hits.total.value = min(k, matches)).
+# * `size` caps the returned hits independently of `k`.
+#
+# The three strings below mirror xerj-ai's `paraphrase_ranks_above_unrelated`
+# unit test, which proves sim(query, doc-1) > sim(query, doc-2) at 384 dims.
+
+setup:
+ - do:
+ indices.create:
+ index: semantic-builtin
+ body:
+ mappings:
+ properties:
+ body:
+ type: semantic_text
+
+ - do:
+ index:
+ index: semantic-builtin
+ id: "1"
+ refresh: true
+ body:
+ body: "a hungry dog chased the ball across the green park"
+
+ - do:
+ index:
+ index: semantic-builtin
+ id: "2"
+ refresh: true
+ body:
+ body: "investors sold technology shares amid rising interest rates"
+
+---
+"semantic query over a semantic_text field ranks the vocabulary-overlapping doc first":
+ - do:
+ search:
+ index: semantic-builtin
+ body:
+ size: 1
+ query:
+ semantic:
+ field: body
+ query: "a dog ran after a ball in the park"
+ k: 10
+
+ # k=10 bounds the neighbour pool; only 2 docs carry a vector → total is 2.
+ - match: { hits.total.value: 2 }
+ # size:1 caps the returned hits regardless of k.
+ - length: { hits.hits: 1 }
+ # The vocabulary-overlapping doc (id 1) must rank first.
+ - match: { hits.hits.0._id: "1" }
+
+---
+"size 0 semantic query returns the pool total but no hits":
+ - do:
+ search:
+ index: semantic-builtin
+ body:
+ size: 0
+ query:
+ semantic:
+ field: body
+ query: "a dog ran after a ball in the park"
+ k: 10
+
+ - match: { hits.total.value: 2 }
+ - length: { hits.hits: 0 }