diff --git a/include/ircd/gpt/token.h b/include/ircd/gpt/token.h
index b928c8758..e7b953468 100644
--- a/include/ircd/gpt/token.h
+++ b/include/ircd/gpt/token.h
@@ -29,7 +29,7 @@ class ircd::gpt::token
 	operator string_view() const;
 
 	token(const_buffer &buf) noexcept;
-	token(const string_view &);
+	token(const string_view &, const bool prefix_space = false);
 	token(const uint16_t &) noexcept;
 };
 
@@ -44,9 +44,11 @@ noexcept
 {}
 
 /// Must resolve to one token or error thrown.
+/// prefix_space=true internally prepends space for potentially better token.
 inline
-ircd::gpt::token::token(const string_view &str)
-:val{vocab::tokenize(str)}
+ircd::gpt::token::token(const string_view &str,
+                        const bool prefix_space)
+:val{vocab::tokenize(str, prefix_space)}
 {}
 
 /// Consumes input for one token off front of buf
diff --git a/include/ircd/gpt/vocab.h b/include/ircd/gpt/vocab.h
index ac314d797..c63d8637e 100644
--- a/include/ircd/gpt/vocab.h
+++ b/include/ircd/gpt/vocab.h
@@ -40,7 +40,7 @@ namespace ircd::gpt::vocab
 	u16 tokenize(const_buffer &) noexcept;
 
 	// Tokenize one token. Error thrown if input is not exactly one token.
-	u16 tokenize(const string_view &in);
+	u16 tokenize(const string_view &in, const bool prefix_space = false);
 
 	// Decode token values to build output text string.
 	string_view detokenize(const mutable_buffer &out, const vector_view<const u16> &in) noexcept;
diff --git a/ircd/gpt_vocab.cc b/ircd/gpt_vocab.cc
index 282059be1..a53c2b4c5 100644
--- a/ircd/gpt_vocab.cc
+++ b/ircd/gpt_vocab.cc
@@ -257,18 +257,24 @@ noexcept
 //
 
 uint16_t
-ircd::gpt::vocab::tokenize(const string_view &in)
+ircd::gpt::vocab::tokenize(const string_view &in,
+                           const bool prefix_space)
 {
-	char str_buf[16];
-	const string_view str
+	char str_buf[16] {' '};
+	const mutable_buffer buf
 	{
-		str_buf, copy(str_buf, in)
+		str_buf + prefix_space, sizeof(str_buf) - prefix_space
 	};
 
-	u16 buf[16];
+	const string_view str
+	{
+		str_buf, copy(buf, in) + prefix_space
+	};
+
+	u16 out_buf[16];
 	const auto out
 	{
-		tokenize(buf, str)
+		tokenize(out_buf, str)
 	};
 
 	if(unlikely(out.size() != 1))
@@ -278,7 +284,7 @@ ircd::gpt::vocab::tokenize(const string_view &in)
 			out.size()
 		};
 
-	return buf[0];
+	return out_buf[0];
 }
 
 uint16_t