feat: port template upgrades (citations, eval, chunking, nomic prefixes)

mcp 2.x is already on origin/main. This ports the rest of
docs-mcp-template #12–#16 without replacing HVM's BM25-first
search_docs path.

- Numbered [1] citations via docs_mcp/format.py
- Eval P@1 + JSONL sidecar + eval.pvalue
- Heading-recursive chunker, keep chunk-0 and MAX_CHARS=4000
- Nomic prefixes at embed time only; stored text unprefixed

Closes #15

rag/embeddings.py

@@ -41,6 +41,21 @@ def _resolve_urls() -> list[str]:
 OLLAMA_URLS = _resolve_urls()
 EMBED_MODEL = os.environ.get("EMBED_MODEL", "nomic-embed-text")
 EMBED_DIM = int(os.environ.get("EMBED_DIM", "768"))
+# nomic asymmetric prefixes — embed time only. Empty string disables.
+EMBED_QUERY_PREFIX = os.environ.get("EMBED_QUERY_PREFIX", "search_query: ")
+EMBED_DOC_PREFIX = os.environ.get("EMBED_DOC_PREFIX", "search_document: ")
+
+
+def apply_prefix(text: str, prefix: str) -> str:
+    if not prefix:
+        return text
+    return prefix + text
+
+
+def embed_texts(texts: list[str], *, prefix: str, ef: Any | None = None) -> list[list[float]]:
+    """Embed texts after applying prefix. Callers store the original strings."""
+    fn = ef if ef is not None else embedding_function()
+    return list(fn([apply_prefix(t, prefix) for t in texts]))
 
 
 class OllamaEmbeddings(EmbeddingFunction):