Some kind of tokenizing, accidentally started creating a taxonomy instead of solving the problem

makeprompt.h

@@ -94,7 +94,6 @@ typedef struct TextChunkList
 	TextChunk text;
 } TextChunkList;
 
-
 typedef enum
 {
 	ARG_CHARACTER,
@@ -286,11 +285,16 @@ typedef enum HealthStatus {
 	HEALTH_verge_of_death,
 } HealthStatus;
 
+// Whatever this health is, it can be perceived by others, e.g it's public
 typedef struct Health {
 	HealthStatus status;
 	float drunkenness; // 1.0 is max drunkenness
 } Health;
 
+
+// these are items and events that are available during the game, but 'rendered' to different structs
+// when sent to the AI as text so that they're more stable. I.E, if you change the name of an item or an index,
+// old memories still work, old reactions to items in a room still work, etc. 
 typedef enum ItemKind {
 	ITEM_none,
 	ITEM_whiskey,
@@ -307,36 +311,84 @@ typedef enum EventKind {
 	EVENT_stopped_talking,
 } EventKind;
 
+// these are the structs as presented to the AI, without a dependence on game data.
+
+typedef struct ItemInSituation {
+	TextChunk name;
+	TextChunk description; // might include some state, e.g 'the beer was drank 5 times'
+	ItemKind actual_item_kind; // used to map back to gameplay items, sometimes might be invalid item if that's not the purpose of the item in a situation
+} ItemInSituation;
+
 typedef struct Response {
-	TextChunk text;
-	ActionKind action;
-	NpcKind target;
-	int memory_slot;
+	TextChunk text; // for speech or memory
+	
+	// both of these indices correspond to what was provided in the CharacterSituation
+	int action_index;
+	int target_index;
+
+	int memory_index;
 } Response;
 
 typedef BUFF(Response, 5) FullResponse; // what the AI is allowed to output 
 
 typedef struct CharacterPerception {
 	Health health;
+	TextChunk name;
+	ItemInSituation held_item;
 } CharacterPerception;
 
+typedef struct ScenePerception {
+	BUFF(CharacterPerception, 10) characters;
+	BUFF(ItemKind, 10) items_on_floor; // available to be picked up or navigated to
+	
+} ScenePerception;
+
 typedef struct CharacterStatus {
-	TextChunk goal;
-	u64 in_room;
 	Item held_item;
 	Health health;
 } CharacterStatus;
 
+typedef enum TargetKind {
+	TARGET_invalid,
+	TARGET_path,
+	TARGET_person,
+	TARGET_item,
+} TargetKind;
+
+typedef struct Target {
+	TextChunk name;
+	TextChunk description;
+	TargetKind kind;
+} Target;
+
 // the situation for somebody
 typedef struct CharacterSituation {
+	TextChunk goal; // kind of like the most important memory, self described character's goal right now
 	TextChunk memories[4]; // explicit numbered memories
-	BUFF(TextChunk, 5) events; // events that this character has observed
+	BUFF(TextChunk, 5) events; // events that this character has observed in the plain english form
+	BUFF(Target, 10) targets;
+	CharacterStatus my_status;
+
 	CharacterStatus status;
 } CharacterSituation;
 
+/*
+Training samples must remain stable as the game is changed, is the decision here: i.e, if the characters
+in the situations are edited/changed, the training samples KEEP the old characters. This is so custom characters
+don't become invalid when the game is updated. I'm making the same decision with the items, instead of storing
+an across-update-stable 'item ID', I want to store the name of the item. All items can be used and have the same 'API'
+so no need of updating there is necessary.
+
+I.E: the situation is freestanding and doesn't refer to any other data. Not NpcKind, not ItemKind, not anything.
+Even the list of available actions and their arguments are stored in the situation. It's basically like pure JSON data.
+
+Also, you can't deserialize training samples easily because they exact text, and such a thing should NEVER happen.
+Training sample into gamestate = bad time. For things like recording gamestate for replays, there's a real man serialization
+codepath into binary.
+*/
 typedef struct TrainingSample {
 	CharacterSituation situation;
-	Response response;
+	FullResponse response;
 } TrainingSample;
 
 typedef struct Npc {

playground.c

@@ -29,41 +29,33 @@
 
 #include <windows.h> // for sleep.
 
-void error_impl(Arena *arena, String8List *errors, String8 message)
+typedef struct Error {
+    struct Error *next, *prev;
+    String8 message;
+    int line;
+} Error;
+
+typedef struct ErrorList {
+    int count;
+    Error *first, *last;
+} ErrorList;
+
+void ErrorPush(Arena *arena, ErrorList *list, Error message)
 {
-    S8ListPush(arena, errors, message);
-    // this is a function so you can breakpoint here and discover when errors occur
-}
+    Error *new_err = PushArrayZero(arena, Error, 1);
+    *new_err = message;
     
-#define error(fmt_str, ...) error_impl(arena, errors, S8Fmt(arena, fmt_str, __VA_ARGS__))
-
-// Allows you to not need to quote children of a parent.
-String8 all_children_as_string(Arena *arena, Node *parent)
-{
-    String8List children = {0};
-    for (Node *cur = parent->first_child; !NodeIsNil(cur); cur = cur->next)
-    {
-        S8ListPush(arena, &children, cur->string);
-    }
-    return S8ListJoin(arena, children, &(StringJoin){.mid = S8Lit(" ")});
+    QueuePush(list->first, list->last, new_err);
+    list->count += 1;
 }
 
-Node *get_child_should_exist(Arena *arena, String8List *errors, Node *parent, String8 child_name)
+void error_impl(Arena *arena, ErrorList *errors, int line_in_toparse, String8 message)
 {
-    if (errors->node_count > 0)
-        return NilNode();
-
-    Node *child = MD_ChildFromString(parent, child_name, StringMatchFlag_CaseInsensitive);
-    if (NodeIsNil(child))
-    {
-        error("Couldn't find child with name '%.*s' on node '%.*s'", S8VArg(child_name), S8VArg(parent->string));
-    }
-    return child;
+    ErrorPush(arena, errors, (Error){.line = line_in_toparse, .message = message});
+    // this is a function so you can breakpoint here and discover when errors occur
 }
 
-typedef struct {
-
-}
+#define error(line_in_toparse, fmt_str, ...) error_impl(arena, errors, line_in_toparse, S8Fmt(arena, fmt_str, __VA_ARGS__))
 
 int main(int argc, char **argv)
 {
@@ -71,41 +63,72 @@ int main(int argc, char **argv)
     (void)argv;
 
     Arena *arena = ArenaAlloc();
-    ParseResult result = ParseWholeFile(arena, S8Lit("playground.mdesk"));
-    if (result.errors.node_count > 0)
-    {
-        printf("Failed to parse file:\n");
-        for (Message *cur = result.errors.first; cur->next; cur = cur->next)
+    Sleep(200); // have to wait for console to pop up, super annoying
+
+    // Prose is the name for this file format where you describe the souls of characters
+
+    // tokenize
+    typedef struct ProseToken
     {
-            printf("%.*s\n", S8VArg(cur->string));
-        }
-    }
-    else
+        struct ProseToken *next, *prev;
+        String8 field;
+        int field_number; // this is -1 if no field_number, e.g if 'Field Text #0:' isn't specified and had no '#', then this would be -1
+        String8 value; // may be an empty string, if it's trying to be like, an object
+        int indentation;
+        int line;
+    } ProseToken;
+
+    ErrorList errors_lit = {0};
+    ErrorList *errors = &errors_lit;
+
+    Npc out = {0};
+    // all arena allocations done from here are temporary. As it just copies data into Npc
+    // parse 'playground.txt' into 'out'
     {
-        Node *node = result.node;
-        String8List errors_list = {0};
-        String8List *errors = &errors_list;
+        // read the file
+        String8 to_parse = LoadEntireFile(arena, S8Lit("playground.txt"));
 
-        Npc out;
-        chunk_from_s8(&out.name, all_children_as_string(arena, get_child_should_exist(arena, errors, node, S8Lit("name"))));
-
-        if (errors->node_count == 0)
+        // tokenize to_parse
+        ProseToken *tokenized_first = 0;
+        ProseToken *tokenized_last = 0;
+        {
+            String8List as_lines = S8Split(arena, to_parse, 1, &S8Lit("\n"));
+            int line = 1; // lines start at 1
+            for (String8Node *cur = as_lines.first; cur; cur = cur->next)
             {
-            // unit testing asserts
-            assert(S8Match(TextChunkString8(out.name), S8Lit("Roger Penrose"), 0));
+                int indentation = 0;
+                while(indentation < cur->string.size && cur->string.str[indentation] == '\t') indentation += 1;
+
+                String8 no_funny_business = S8SkipWhitespace(S8ChopWhitespace(cur->string));
+                if(no_funny_business.size == 0) continue;
+
+                String8List along_colon = S8Split(arena, no_funny_business, 1, &S8Lit(":"));
+                if(along_colon.node_count != 2 && along_colon.node_count != 1) {
+                    error(line, "Requires exactly one ':' on the line to delimit the field and value. Got %d", along_colon.node_count - 1);
+                } else {
+                    ProseToken *token_out = PushArrayZero(arena, ProseToken, 1);
+                    token_out->field_number = -1;
+                    if(along_colon.node_count == 2)
+                        token_out->value = along_colon.last->string;
+                    token_out->line = line;
+                    token_out->indentation = indentation;
+
+                    DblPushBack(tokenized_first, tokenized_last, token_out);
+                }
+                line += 1;
             }
-        else
+        }
+    }
+
+    if (errors->count > 0)
     {
-            printf("Corrupt character soul:\n");
-            for (String8Node *cur = errors->first; cur->next; cur = cur->next)
+        printf("Failed with errors:\n");
+        for (Error *cur = errors->first; cur; cur = cur->next)
         {
-                printf("%.*s\n", S8VArg(cur->string));
+            printf("On line %d of input: %.*s\n", cur->line, S8VArg(cur->message));
         }
         assert(false);
     }
-
-    }
-
     printf("Success.\n");
     __debugbreak();
 }

playground.txt

@@ -6,6 +6,7 @@ Soul:
 		Memory #1: Silly
 
 		Perception:
+		Some testing thing:
 			In Room: Farm
 			Characters Around Me:
 				John:
@@ -29,3 +30,4 @@ Soul:
 				Speech: Why are you like that dude?
 			Throw:
 				To: John
+