Commit a123d9f

Cleanup and refactor

Author	Mitja Felicijan <mitja.felicijan@gmail.com>	2026-02-18 13:26:22 +0100
Committer	Mitja Felicijan <mitja.felicijan@gmail.com>	2026-02-18 13:26:22 +0100
Commit	`a123d9f1ebebb6ad20155285c9457a6d6a5c7a61` (patch)

`-rw-r--r--`	Makefile	13
`-rw-r--r--`	README.md	26
`-rw-r--r--`	corpus/lotr.txt	0
`-rw-r--r--`	npc.c	8
`-rw-r--r--`	prompts/lotr.h	0
`-rw-r--r--`	prompts/lotr.txt	0

6 files changed, 34 insertions, 17 deletions

 		  -lpthread -lm -ldl -lstdc++ -g \
 		  -lllama -lggml -lggml-cpu -lggml-base
+PROMPT_TXT := $(wildcard prompts/*.txt)
+PROMPT_HEADERS := $(PROMPT_TXT:.txt=.h)
 help: .help
-build/npc: run/system-prompt npc.c vectordb.c models.h # Build npc binary for testing
+build/npc: build/prompts npc.c vectordb.c models.h # Build npc binary for testing
 	$(CC) $(CFLAGS) npc.c vectordb.c -o npc $(LDFLAGS)
 build/context: context.c vectordb.c models.h # Build context binary for testing
 		cd $(LLAMA_DIR)/build && \
 		cmake ../ -DBUILD_SHARED_LIBS=OFF && \
 		make -j8
+build/prompts: $(PROMPT_HEADERS) # Generate C style header
 run/fetch-models: .assure # Fetch GGUF models
 	-mkdir -p models
 	docker build -t npcd .
 	docker run -it npcd
-run/system-prompt: .assure # Generate C style header
-	xxd -i system_prompt.txt > system_prompt.h
 run/clean: # Cleans up all the build artefacts
 	-rm -f npc
 	cd $(LLAMA_DIR)/build && make clean
 	-rm -Rf $(LLAMA_DIR)/build
+prompts/%.h: prompts/%.txt .assure
+	xxd -i $< > $@

-# llmnpc
+An experiment using tiny LLMs as NPCs that could be embedded into the game.
+Goals of the experiment:
-Command-line tooling for NPC-focused LLM experiments with lightweight context
+- Have LLM be run only on CPU, this is why small LLMs have been chosen in this
-retrieval, powered by [llama.cpp](https://github.com/ggerganov/llama.cpp).
+  experiment, so they can be used in other games.
+- To produce a simple C library that can be reused elsewhere.
+- Test existing small and tiny LLMs and provide some useful results on how they
+  behave.
+> [!NOTE]
+> This project is just for fun, to see how LLMs would fare as NPCs. Because of
+> the non-deterministic nature of LLMs, the results vary and are often quite
+> funny. A lot of tweaking would be needed to make this really useful in real
+> games, but not impossible.
 ## Building
 3. Build binaries:
    ```bash
    make build/context
+   make build/prompts
    make build/npc
    ```
 produces a binary vector database file.
 ```bash
-./context -i context.txt -o context.vdb
+./context -i corpus/lotr.txt -o corpus/lotr.vdb
-./context -m flan-t5-small -i context.txt -o context.vdb
+./context -m flan-t5-small -i corpus/lotr.txt -o corpus/lotr.vdb
 ```
 ### Run an NPC query with retrieved context
 lines by cosine similarity, and runs the NPC system prompt against that context.
 ```bash
-./npc -m flan-t5-small -p "Who is Gandalf?" -c context.vdb
+./npc -m flan-t5-small -p "Who is Gandalf?" -c corpus/lotr.vdb
-./npc -m flan-t5-small -p "Who is Frodo?" -c context.vdb
+./npc -m flan-t5-small -p "Who is Frodo?" -c corpus/lotr.vdb
 ```
 ### context options

diff --git a/corpus/lotr.txt b/corpus/lotr.txt
...



#include <string.h>
#include <getopt.h>
  
#include "prompts/lotr.h"
  
static void llama_log_callback(enum ggml_log_level level, const char *text, void *user_data) {
	(void)level;

		return 1;
	}
  
	char *system_prefix = (char *)malloc(prompts_lotr_txt_len + 1);
	if (system_prefix == NULL) {
		log_message(stderr, LOG_ERROR, "Failed to allocate system prompt");
		return 1;
	}
	memcpy(system_prefix, prompts_lotr_txt, prompts_lotr_txt_len);
	system_prefix[prompts_lotr_txt_len] = '\0';
  
	ggml_backend_load_all();
  


diff --git a/prompts/lotr.h b/prompts/lotr.h
1	unsigned char system_prompt_txt[] = {	1	unsigned char prompts_lotr_txt[] = {
2	0x53, 0x79, 0x73, 0x74, 0x65, 0x6d, 0x3a, 0x20, 0x41, 0x6e, 0x73, 0x77,	2	0x53, 0x79, 0x73, 0x74, 0x65, 0x6d, 0x3a, 0x20, 0x41, 0x6e, 0x73, 0x77,
3	0x65, 0x72, 0x20, 0x75, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x6f, 0x6e, 0x6c,	3	0x65, 0x72, 0x20, 0x75, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x6f, 0x6e, 0x6c,
4	0x79, 0x20, 0x74, 0x68, 0x65, 0x20, 0x43, 0x6f, 0x6e, 0x74, 0x65, 0x78,	4	0x79, 0x20, 0x74, 0x68, 0x65, 0x20, 0x43, 0x6f, 0x6e, 0x74, 0x65, 0x78,
...
12	0x74, 0x68, 0x61, 0x74, 0x20, 0x69, 0x6e, 0x66, 0x6f, 0x72, 0x6d, 0x61,	12	0x74, 0x68, 0x61, 0x74, 0x20, 0x69, 0x6e, 0x66, 0x6f, 0x72, 0x6d, 0x61,
13	0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x0a	13	0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x0a
14	};	14	};
15	unsigned int system_prompt_txt_len = 138;	15	unsigned int prompts_lotr_txt_len = 138;

diff --git a/prompts/lotr.txt b/prompts/lotr.txt
...

diff --git a/npc.c b/npc.c
...
11	#include <string.h>	11	#include <string.h>
12	#include <getopt.h>	12	#include <getopt.h>
13		13
14	#include "system_prompt.h"	14	#include "prompts/lotr.h"
15		15
16	static void llama_log_callback(enum ggml_log_level level, const char text, void user_data) {	16	static void llama_log_callback(enum ggml_log_level level, const char text, void user_data) {
17	(void)level;	17	(void)level;
...
54	return 1;	54	return 1;
55	}	55	}
56		56
57	char system_prefix = (char )malloc(system_prompt_txt_len + 1);	57	char system_prefix = (char )malloc(prompts_lotr_txt_len + 1);
58	if (system_prefix == NULL) {	58	if (system_prefix == NULL) {
59	log_message(stderr, LOG_ERROR, "Failed to allocate system prompt");	59	log_message(stderr, LOG_ERROR, "Failed to allocate system prompt");
60	return 1;	60	return 1;
61	}	61	}
62	memcpy(system_prefix, system_prompt_txt, system_prompt_txt_len);	62	memcpy(system_prefix, prompts_lotr_txt, prompts_lotr_txt_len);
63	system_prefix[system_prompt_txt_len] = '\0';	63	system_prefix[prompts_lotr_txt_len] = '\0';
64		64
65	ggml_backend_load_all();	65	ggml_backend_load_all();
66		66
...