fix(cross_ontology_search): embed cache, strict cosine, early qv=None return, hardware comment

[Copilot]

Four independent improvements to metpo/pipeline/cross_ontology_search.py addressing correctness, performance, and clarity.

Changes

• Comment — replace "M5 GPU" with "CPU/GPU-capable hardware"; the original reference was ambiguous/non-standard

• Embedding cache — add module-level _EMBED_CACHE and _get_cached_embedding() to deduplicate Ollama API calls within a run; previously the same text (e.g. a label shared across multiple terms) would trigger a fresh HTTP round-trip each time

  _EMBED_CACHE = {}
  
  def _get_cached_embedding(text, model, embed_url):
      key = (model, embed_url, text.strip(". "))
      if key not in _EMBED_CACHE:
          _EMBED_CACHE[key] = embed(key[2], model, embed_url)
      return _EMBED_CACHE[key]

• Strict cosine() — explicit dimension-mismatch check raises ValueError before the dot product; zip() upgraded to strict=True as a belt-and-suspenders guard; mismatched vectors were previously silently truncated

• Early return on qv = None in rank_candidates()** — when the query embedding fails, fall back immediately to lexical ranking instead of looping over all candidates, calling embed() for each, and assigning misleading 0.0 scores; also drops the now-redundant qv and from the per-candidate guard

Original prompt

Please apply the following diffs and create a pull request.
Once the PR is ready, give it a title based on the messages of the fixes being applied.

[{"message":"The comment mentions 'M5 GPU' which appears to be a typo or non-standard GPU designation. Common GPU models include M1/M2/M3 (Apple Silicon) or NVIDIA/AMD models. Consider clarifying the intended GPU type or removing the specific hardware reference if it's meant to be generic.","fixFiles":[{"filePath":"metpo/pipeline/cross_ontology_search.py","diff":"diff --git a/metpo/pipeline/cross_ontology_search.py b/metpo/pipeline/cross_ontology_search.py\n--- a/metpo/pipeline/cross_ontology_search.py\n+++ b/metpo/pipeline/cross_ontology_search.py\n@@ -36,7 +36,7 @@\n import click\n \n OLS_SEARCH = \"https://www.ebi.ac.uk/ols4/api/search\"\n-# Default embedding backend: free, local, runs on the M5 GPU. Nothing paid, no API\n+# Default embedding backend: free, local, runs on CPU/GPU-capable hardware. Nothing paid, no API\n # key. Both the endpoint and the model are CLI-parameterized so anyone can point at\n # a different local model, a remote Ollama server, or other hardware.\n DEFAULT_EMBED_URL = \"http://localhost:11434/api/embeddings\"\n"}]},{"message":"The `strict=False` parameter in `zip()` silently truncates if vectors have different lengths, which could mask data quality issues. For cosine similarity calculations, vectors should have matching dimensions. Consider using `strict=True` to catch dimension mismatches early, or add explicit length validation before the calculation.","fixFiles":[{"filePath":"metpo/pipeline/cross_ontology_search.py","diff":"diff --git a/metpo/pipeline/cross_ontology_search.py b/metpo/pipeline/cross_ontology_search.py\n--- a/metpo/pipeline/cross_ontology_search.py\n+++ b/metpo/pipeline/cross_ontology_search.py\n@@ -91,7 +91,10 @@\n \n \n def cosine(a, b):\n-    dot = sum(x * y for x, y in zip(a, b, strict=False))\n+    if len(a) != len(b):\n+        msg = f\"cosine vector dimension mismatch: len(a)={len(a)} != len(b)={len(b)}\"\n+        raise ValueError(msg)\n+    dot = sum(x * y for x, y in zip(a, b, strict=True))\n     na = math.sqrt(sum(x * x for x in a))\n     nb = math.sqrt(sum(x * x for x in b))\n     return dot / (na * nb) if na and nb else 0.0\n"}]},{"message":"The function makes N+1 embedding API calls (1 for query + N for candidates), which can be slow for large candidate sets. Consider batching candidate embeddings if the Ollama API supports batch requests, or caching embeddings for frequently-queried candidates to reduce redundant API calls.","fixFiles":[{"filePath":"metpo/pipeline/cross_ontology_search.py","diff":"diff --git a/metpo/pipeline/cross_ontology_search.py b/metpo/pipeline/cross_ontology_search.py\n--- a/metpo/pipeline/cross_ontology_search.py\n+++ b/metpo/pipeline/cross_ontology_search.py\n@@ -42,7 +42,10 @@\n DEFAULT_EMBED_URL = \"http://localhost:11434/api/embeddings\"\n DEFAULT_EMBED_MODEL = \"nomic-embed-text\"\n \n+# In-memory cache to avoid redundant embedding API calls within a run.\n+_EMBED_CACHE = {}\n \n+\n def ols_candidates(label, rows=20):\n     \"\"\"Lexical candidate classes from OLS4 for a label (IRI, curie, label, def, ontology).\"\"\"\n     q = urllib.parse.urlencode({\"q\": label, \"type\": \"class\", \"rows\": rows})\n@@ -90,6 +92,14 @@\n         return None\n \n \n+def _get_cached_embedding(text, model, embed_url):\n+    normalized = text.strip(\". \")\n+    key = (model, embed_url, normalized)\n+    if key not in _EMBED_CACHE:\n+        _EMBED_CACHE[key] = embed(normalized, model, embed_url)\n+    return _EMBED_CACHE[key]\n+\n+\n def cosine(a, b):\n     dot = sum(x * y for x, y in zip(a, b, strict=False))\n     na = math.sqrt(sum(x * x for x in a))\n@@ -130,9 +140,9 @@\n def rank_candidates(term, cands, use_embeddings, model, embed_url):\n     \"\"\"Attach a 'similarity' score to each candidate (embedding cosine or lexical order).\"\"\"\n     if use_embeddings and cands:\n-        qv = embed((term[\"label\"] + \". \" + term[\"definition\"]).strip(\". \"), model, embed_url)\n+        qv = _get_cached_embedding(term[\"label\"] + \". \" + term[\"definition\"], model, embed_url)\n         for c in cands:\n-            cv = embed((c[\"label\"] + \". \" + c[\"definition\"]).strip(\". \"), model, embed_url)\n+            cv = _get_cached_embedding(c[\"label\"] + \". \" + c[\"definition\"], model, embed_url)\n             c[\"similarity\"] = round(cosine(qv, cv), 4) if (qv and cv) else 0.0\n         cands.sort(key=lambda c: c[\"similarity\"], reverse=True)\n     else:\n"}]},{"message":"If `qv` (query vector) is None, the function continues to embed all candidates but assigns similarity=0.0 to all. This wastes API calls and produces misleading zero-similarity scores. Consider returning early or skipping embedding calls when qv is None to avoid unnecessary work and confusing output.","fixFiles":[{"filePath":"metpo/pipeline/cross_ontology_search.py","diff":"diff --git a/metpo/pi...

[Copilot]

Pull request overview

This PR improves metpo/pipeline/cross_ontology_search.py by reducing redundant embedding calls, tightening cosine similarity correctness, and improving behavior when embedding is unavailable, while also clarifying a hardware-related comment.

Changes:

• Add an in-process embedding cache to deduplicate repeated Ollama /api/embeddings requests within a run.
• Make cosine() strict about vector dimension mismatches (explicit check + zip(..., strict=True)).
• If the query embedding fails (qv is None), immediately fall back to lexical ranking instead of embedding every candidate and assigning misleading 0.0 similarities.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.