added support for structured output in chat completions

koboldcpp.py

@@ -1267,6 +1267,14 @@ def generate(genparams, stream_flag=False):
     ban_eos_token = genparams.get('ban_eos_token', False)
     stream_sse = stream_flag
     grammar = genparams.get('grammar', '')
+    #translate grammar if its json
+    try:
+        grammarjson = json.loads(grammar)
+        decoded = convert_json_to_gbnf(grammarjson)
+        if decoded:
+            grammar = decoded
+    except Exception:
+        pass
     grammar_retain_state = genparams.get('grammar_retain_state', False)
     genkey = genparams.get('genkey', '')
     trimstop = genparams.get('trim_stop', True)
@@ -2051,6 +2059,32 @@ def transform_genparams(genparams, api_format):
             tools_message_start = adapter_obj.get("tools_start", "")
             tools_message_end = adapter_obj.get("tools_end", "")
             images_added = []
+            jsongrammar = r"""
+root   ::= arr
+value  ::= object | array | string | number | ("true" | "false" | "null") ws
+arr  ::=
+  "[\n" ws (
+            value
+    (",\n" ws value)*
+  )? "]"
+object ::=
+  "{" ws (
+            string ":" ws value
+    ("," ws string ":" ws value)*
+  )? "}" ws
+array  ::=
+  "[" ws (
+            value
+    ("," ws value)*
+  )? "]" ws
+string ::=
+  "\"" (
+    [^"\\\x7F\x00-\x1F] |
+    "\\" (["\\bfnrt] | "u" [0-9a-fA-F]{4})
+  )* "\"" ws
+number ::= ("-"? ([0-9] | [1-9] [0-9]{0,15})) ("." [0-9]+)? ([eE] [-+]? [1-9] [0-9]{0,15})? ws
+ws ::= | " " | "\n" [ \t]{0,20}
+"""
 
             # tools handling
             tools_array = genparams.get('tools', [])
@@ -2075,6 +2109,24 @@ def transform_genparams(genparams, api_format):
                         tool_json_formatting_instruction = f"\nThe user is asking you to use the style of this JSON object formatting to complete the parameters for the specific function named {specified_function} in the following format: " + json.dumps([{"id": "insert an id for the response", "type": "function", "function": {"name": f"{specified_function}", "arguments": {"first property key": "first property value", "second property key": "second property value"}}}], indent=0)
                 except Exception:
                     # In case of any issues, just revert back to no specified function
+                    print("Tools parsing not valid - discarded")
+                    pass
+
+            # handle structured outputs
+            respformat = genparams.get('response_format', None)
+            if respformat:
+                try:
+                    rt = respformat.get('type')
+                    if rt.lower() == "json_schema":
+                        schema = respformat.get('json_schema').get('schema')
+                        decoded = convert_json_to_gbnf(schema)
+                        if decoded:
+                            genparams["grammar"] = decoded
+                    elif rt.lower() == "json_object":
+                        genparams["grammar"] = jsongrammar
+                except Exception:
+                    # In case of any issues, just do normal gen
+                    print("Structured Output not valid - discarded")
                     pass
 
             message_index = 0
@@ -2115,13 +2167,15 @@ def transform_genparams(genparams, api_format):
                             # if you want a different template, you can set 'custom_tools_prompt' in the chat completions adapter as follows
                             custom_tools_prompt = adapter_obj.get("custom_tools_prompt", "Can the user query be answered by a listed tool? (One word response: yes or no):")
                             # note: message string already contains the instruct start tag!
+                            pollgrammar = r'root ::= "yes" | "no" | "Yes" | "No" | "YES" | "NO"'
                             temp_poll = {
                                 "prompt": f"{messages_string}\n\nTool List:\n{tools_string}\n\n{custom_tools_prompt}{user_end}",
-                                "max_length":6,
+                                "max_length":4,
                                 "temperature":0.1,
                                 "top_k":1,
                                 "rep_pen":1,
-                                "ban_eos_token":False
+                                "ban_eos_token":False,
+                                "grammar":pollgrammar
                                 }
                             temp_poll_result = generate(genparams=temp_poll)
                             if temp_poll_result and "yes" not in temp_poll_result['text'].lower():
@@ -2138,32 +2192,7 @@ def transform_genparams(genparams, api_format):
                             genparams["using_openai_tools"] = True
 
                             # Set grammar to llamacpp example grammar to force json response (see https://github.com/ggerganov/llama.cpp/blob/master/grammars/json_arr.gbnf)
-                            genparams["grammar"] = r"""
-root   ::= arr
-value  ::= object | array | string | number | ("true" | "false" | "null") ws
-arr  ::=
-  "[\n" ws (
-            value
-    (",\n" ws value)*
-  )? "]"
-object ::=
-  "{" ws (
-            string ":" ws value
-    ("," ws string ":" ws value)*
-  )? "}" ws
-array  ::=
-  "[" ws (
-            value
-    ("," ws value)*
-  )? "]" ws
-string ::=
-  "\"" (
-    [^"\\\x7F\x00-\x1F] |
-    "\\" (["\\bfnrt] | "u" [0-9a-fA-F]{4})
-  )* "\"" ws
-number ::= ("-"? ([0-9] | [1-9] [0-9]{0,15})) ("." [0-9]+)? ([eE] [-+]? [1-9] [0-9]{0,15})? ws
-ws ::= | " " | "\n" [ \t]{0,20}
-"""
+                            genparams["grammar"] = jsongrammar
                 if message['role'] == "system":
                     messages_string += system_message_end
                 elif message['role'] == "user":