diff --git a/README.md b/README.md
index 434d642..eea6321 100644
--- a/README.md
+++ b/README.md
@@ -154,7 +154,8 @@ each row must be judged in complete isolation.
   applied before inference, so `WHERE region = 'eu' AND prompt_jev(...) > 0.8`
   only asks about the `eu` rows.
 - **Limits.** Each row's text may be up to 64 KiB after JSON encoding. Longer text
-  fails the query. Requests are capped at 256 KiB.
+  fails the query. Requests are capped at 256 KiB. Questions may be up to 4,000
+  characters, labels 256, and descriptions 1,024.
 - **Outages.** Each request is tried three times with a 30-second timeout. If all
   fail, the affected rows return NULL and the query completes. Bad credentials,
   invalid options, or a malformed reply fail the query instead.
diff --git a/src/sql.rs b/src/sql.rs
index f8a2913..54d32c4 100644
--- a/src/sql.rs
+++ b/src/sql.rs
@@ -21,12 +21,25 @@ pub struct Question {
     pub batch_size: usize,
 }
 impl Question {
+    /// Longest instructions accepted, in characters. The provider rejects
+    /// requests well before the crate's request-size cap would, and a plan-time
+    /// error names the actual problem instead of a provider HTTP status.
+    pub const MAX_INSTRUCTION_CHARS: usize = 4_000;
+    pub const MAX_LABEL_CHARS: usize = 256;
+    pub const MAX_DESCRIPTION_CHARS: usize = 1_024;
+
     pub fn validate(&self) -> Result<()> {
         if self.instructions.trim().is_empty() {
             return Err(plan_datafusion_err!(
                 "prompt_jev instructions must not be empty"
             ));
         }
+        if self.instructions.chars().count() > Self::MAX_INSTRUCTION_CHARS {
+            return Err(plan_datafusion_err!(
+                "prompt_jev instructions exceed {} characters",
+                Self::MAX_INSTRUCTION_CHARS
+            ));
+        }
         if !(1..=64).contains(&self.batch_size) {
             return Err(plan_datafusion_err!(
                 "prompt_jev batch_size must be between 1 and 64"
@@ -42,6 +55,17 @@ impl Question {
                     "prompt_jev criteria require unique non-empty labels and non-empty descriptions"
                 ));
             }
+            if c.label.chars().count() > Self::MAX_LABEL_CHARS
+                || c.description
+                    .as_ref()
+                    .is_some_and(|d| d.chars().count() > Self::MAX_DESCRIPTION_CHARS)
+            {
+                return Err(plan_datafusion_err!(
+                    "prompt_jev labels are limited to {} characters and descriptions to {}",
+                    Self::MAX_LABEL_CHARS,
+                    Self::MAX_DESCRIPTION_CHARS
+                ));
+            }
         }
         let valid = match self.kind.as_str() {
             "choice" => (2..=255).contains(&self.criteria.len()),
@@ -122,7 +146,15 @@ pub fn rewrite_statement(statement: &mut ast::Statement) -> Result<()> {
         let Expr::Function(f) = expr else {
             return ControlFlow::Continue(());
         };
-        if f.name.to_string().to_lowercase() != "prompt_jev" {
+        // Match the bare identifier, so `"prompt_jev"(...)` and `PROMPT_JEV(...)`
+        // are the same call; a schema-qualified name is left to DataFusion.
+        let is_prompt_jev = match f.name.0.as_slice() {
+            [part] => part
+                .as_ident()
+                .is_some_and(|i| i.value.eq_ignore_ascii_case("prompt_jev")),
+            _ => false,
+        };
+        if !is_prompt_jev {
             return ControlFlow::Continue(());
         }
         let result = (|| {
diff --git a/tests/sql.rs b/tests/sql.rs
index b55b2a0..9b84ce8 100644
--- a/tests/sql.rs
+++ b/tests/sql.rs
@@ -850,3 +850,44 @@ async fn a_folded_limit_survives_the_filter_split() {
         "rows failing `id > 1` must not be asked about: {asked}"
     );
 }
+
+#[tokio::test]
+async fn quoted_and_uppercase_names_are_the_same_function() {
+    let mock = Arc::new(Mock::default());
+    let result = run(
+        mock.clone(),
+        "SELECT \"prompt_jev\"('a', 'q') AS a, PROMPT_JEV('a', 'q') AS b, \"PROMPT_JEV\"('a', 'q') AS c",
+    )
+    .await
+    .unwrap();
+    assert_eq!(result.iter().map(|b| b.num_rows()).sum::<usize>(), 1);
+    assert!(display(&result).contains("0.9"));
+}
+
+#[tokio::test]
+async fn oversized_instructions_and_labels_fail_at_plan_time() {
+    let long_q = "q".repeat(5_000);
+    let long_label = "L".repeat(300);
+    let long_desc = "d".repeat(2_000);
+    for (sql, needle) in [
+        (
+            format!("SELECT prompt_jev('a', '{long_q}')"),
+            "4000 characters",
+        ),
+        (
+            format!("SELECT prompt_jev('a', 'q', choice := ['{long_label}', 'b'])"),
+            "labels are limited",
+        ),
+        (
+            format!(
+                "SELECT prompt_jev('a', 'q', choice := [{{label: 'a', description: '{long_desc}'}}, 'b'])"
+            ),
+            "descriptions to 1024",
+        ),
+    ] {
+        let mock = Arc::new(Mock::default());
+        let err = run(mock.clone(), &sql).await.err().unwrap().to_string();
+        assert!(err.contains(needle), "{needle}: {err}");
+        assert!(mock.calls.lock().unwrap().is_empty());
+    }
+}