Ollama

.gitignore

@@ -1,5 +1,6 @@
 # OpenAI API key
 openai_config.json
+llm_config.json
 
 environment/frontend_server/storage/*
 environment/frontend_server/temp_storage*/*
@@ -115,4 +116,4 @@ ENV/
 .spyderproject
 
 # Rope project settings
-.ropeproject
+.ropeproject

README.md

@@ -13,7 +13,7 @@ Since the project is no longer officially supported, I decided to develop some n
 - [x] **Cost tracking** using [openai-cost-logger](https://github.com/drudilorenzo/openai-cost-logger)
 - [x] Set **cost upperbound** and stop the experiment when it is reached
 - [x] New models and OpenAI API support
-- [x] Added [skip-morning-s-14](https://github.com/drudilorenzo/generative_agents/tree/fix-and-improve/environment/frontend_server/storage/skip-morning-s-14): a simulation based on `base_the_ville_n25` that starts after 3000 steps (~8:00am). That permits us to save time and see interactions and actions earlier.
+- [x] Added [skip-morning-s-14](https://github.com/drudilorenzo/generative_agents/tree/fix-and-improve/environment/frontend_server/storage/skip-morning-s-14): a simulation based on `base_the_ville_n25` that starts after 3000 steps (~8:00am). That permits us to save time and see interactions and actions earlier. (Note that one must use the same embedding model to load this simulation)
 - [x] **Zoom in**/**Zoom out** using Z and X
 - [x] [Powerful automated script](#step-3-automatic-execution) for enhanced simulation performance.
 
@@ -38,7 +38,7 @@ Do not change the env name to be able to use the bash scripts later.
 
 ### Step 2. OpenAI Config
 
-Create a file called `openai_config.json` in the root directory.\
+Create a file called `llm_config.json` in the root directory.\
 Azure example:
 ```json
 {
@@ -85,10 +85,37 @@ OpenAI example:
     "cost-upperbound": 10
 }
 ```
+Ollama example:
+```json
+{
+    "client": "ollama", 
+    "base_url": "http://localhost:11434/v1",
+    "model": "llama3-chatqa:8b",
+    "prompt_template": "prompts_llama3",
+    "model-key": "dummy_key_needed_but_not_used",
+    "model-costs": {
+        "input":  0.001,
+        "output": 0.001
+    },
+    "embeddings-client": "ollama",
+    "embeddings": "nomic-embed-text",
+    "embeddings-key": "dummy_key_needed_but_not_used",
+    "embeddings-costs": {
+        "input":  0.0001,
+        "output": 0.0001
+    },
+    "experiment-name": "simulacra-test-ollama",
+    "cost-upperbound": 10000000000000
+}
+```
 
 Feel free to change and test also other models (and change accordingly the input and output costs).\
-Be aware that the only supported clients are **azure** and **openai**.\
+Be aware that the only fully-supported clients are **azure** and **openai**.\
+The ollama client supports local models, but many of the prompts do not full work with local models yet.
+Make sure to first install ollama, and to pull any models you plan on using, see "[ollama-openai-compatibility](https://ollama.com/blog/openai-compatibility)"\
 The generation and the embedding models are configured separately to be able to use different clients.\
+The llama3 models require different prompts than the chatgpt, and these different prompts are specified via the template_dir argument.
+If you are testing a new model you can use either the prompts_chagpt or the prompts_llama3 as starting points.
 Change also the `cost-upperbound` according to your needs (the cost computation is done using "[openai-cost-logger](https://github.com/drudilorenzo/openai-cost-logger)" and the costs are specified per million tokens).
 
 

README_origin.md

@@ -30,6 +30,8 @@ collision_block_id = "32125"
 
 # Verbose 
 debug = True
+EXCEPT_ON_FAILSAFE=False
+
 ```
 Replace `<Your OpenAI API>` with your OpenAI API key, and `<name>` with your name.
 

log_parser.py

@@ -0,0 +1,147 @@
+import os
+import json
+import re
+
+def parse_section(section, verbose=False):
+    """ Parse the complete section to extract necessary details with an optional verbose output """
+    if verbose:
+        print("Section preview:", section[:200])  # Print beginning of the section to inspect
+
+    repeat_count_match = re.search(r'---- repeat count: \s*(\d+)', section)
+    repeat_count = int(repeat_count_match.group(1)) if repeat_count_match else 0
+
+    prompt_inputs = re.findall(r'---- prompt_input_\d+ (.*?)(?=----|$)', section, re.DOTALL)
+
+    prompt = re.findall(r'---- prompt: (.*?)(?=----|$)', section, re.DOTALL)[0]
+
+    func_validate = re.findall(r'---- func_validate: (.*?)(?=----|$)', section, re.DOTALL)[0]
+
+    gpt_parameter = re.findall(r'---- gpt_parameter: (.*?)(?=----|$)', section, re.DOTALL)[0]
+
+    prompt_template = re.findall(r'---- prompt_template: (.*?)(?=----|$)', section, re.DOTALL)[0].strip()
+
+    curr_response = re.findall(r'---- curr_gpt_response: (.*?)(?=----|$)', section, re.DOTALL)[0]
+
+    func_cleanup = re.findall(r'----  func_clean_up: (.*?)(?=----|$)', section, re.DOTALL)[0]
+
+    func_validate_match = re.search(r'---- func_validate: \s*(True|False)', section)
+    func_validate = func_validate_match.group(1).lower() == 'true' if func_validate_match else False
+
+
+    return {
+        "prompt_inputs": prompt_inputs,
+        "repeat_count": repeat_count,
+        "prompt": prompt,
+        "prompt_template": prompt_template,
+        "gpt_parameter": gpt_parameter,
+        "func_validate": func_validate,
+        "curr_response": curr_response,
+        "func_cleanup": func_cleanup
+    }
+
+#def read_large_file(input_file_path, output_directory, verbose=False):
+#    os.makedirs(output_directory, exist_ok=True)  # Ensure the output directory exists
+#    section = ""
+#    file_count = 0  # To keep track of how many files we have created
+#    fields_found = set()
+#
+#    required_fields = {
+#        "prompt_template_prompt_inputs",
+#        "repeat_count",
+#        "func_validate",
+#        "curr_gpt_response",
+#        "func_cleanup"
+#    }
+#
+#    with open(input_file_path, 'r') as file:
+#        for line in file:
+#            section += line
+#
+#            # Check if the current line contains any of the required fields
+#            if "prompt_input_" in line:
+#                fields_found.add("prompt_template_prompt_inputs")
+#            if "repeat count:" in line:
+#                fields_found.add("repeat_count")
+#            if "func_validate" in line:
+#                fields_found.add("func_validate")
+#            if "curr_gpt_response:" in line:
+#                fields_found.add("curr_gpt_response")
+#            if "func_clean_up:" in line:
+#                fields_found.add("func_cleanup")
+#
+#            # Process the section if all fields are found or if the section delimiter is encountered
+#            if fields_found == required_fields or '~~~~' in line:
+#                if section.strip():  # Make sure there's something to process
+#                    try:
+#                      result = parse_section(section, verbose)
+#                    except:
+#                      raise Exception(section)
+#                    if result:  # Ensure there is a result to write
+#                        json_filename = os.path.join(output_directory, f"{file_count + 1}.json")
+#                        with open(json_filename, 'w') as json_file:
+#                            json.dump(result, json_file, indent=4)
+#                        file_count += 1
+#                # Reset the section and fields found for the next part
+#                section = ""
+#                fields_found.clear()
+#
+#    # Handle the last section if the file does not end with a delimiter
+#    if section.strip() and fields_found == required_fields:
+#        result = parse_section(section, verbose)
+#        if result:
+#            json_filename = os.path.join(output_directory, f"{file_count + 1}.json")
+#            with open(json_filename, 'w') as json_file:
+#                json.dump(result, json_file, indent=4)
+
+def read_large_file(input_file_path, output_directory, verbose=False):
+    os.makedirs(output_directory, exist_ok=True)
+    section = ""
+    file_count = 0
+    in_section = False
+    count_dict = {}  # Dictionary to store file counts per directory
+
+    with open(input_file_path, 'r') as file:
+        for line in file:
+            if "------- BEGIN SAFE GENERATE --------" in line:
+                in_section = True
+                section = line
+            elif "------- END TRIAL" in line and in_section:
+                section += line
+                result = parse_section(section, verbose)
+                if result:
+                    # Extracting the basename from the 'prompt_template' key
+                    template_path = result.get('prompt_template', 'ERROR')
+                    basename = os.path.basename(template_path)
+                    # Removing file extension
+                    basename = os.path.splitext(basename)[0]
+
+                    # Directory for the basename
+                    output_subdir = os.path.join(output_directory, basename)
+                    os.makedirs(output_subdir, exist_ok=True)
+
+                    # Check if we already have a count for this directory, otherwise count files
+                    if basename not in count_dict:
+                        existing_files = [f for f in os.listdir(output_subdir) if f.endswith('.json')]
+                        count_dict[basename] = len(existing_files) + 1
+                    else:
+                        count_dict[basename] += 1
+
+                    # Writing the result to the JSON file
+                    json_filename = os.path.join(output_subdir, f"{count_dict[basename]}.json")
+                    with open(json_filename, 'w') as json_file:
+                        json.dump(result, json_file, indent=4)
+                    file_count += 1
+                section = ""
+                in_section = False
+            elif in_section:
+                section += line
+
+if __name__ == "__main__":
+    import sys
+    if len(sys.argv) < 3:
+        print("Usage: python script.py <input_log_file_path> <output_directory_path> [<verbose>]")
+    else:
+        input_log_file_path = sys.argv[1]
+        output_directory_path = sys.argv[2]
+        verbose = sys.argv[3].lower() == 'true' if len(sys.argv) > 3 else False
+        read_large_file(input_log_file_path, output_directory_path, verbose)

requirements.txt

@@ -66,4 +66,4 @@ wsproto==1.2.0
 yarl==1.8.2
 yellowbrick==1.5
 zipp==3.6.0
-openai-cost-logger==0.4.1
+openai-cost-logger==0.4.1ollama==0.2.1

reverie/backend_server/automatic_execution.py

@@ -15,6 +15,8 @@
 from multiprocessing import Process
 from openai_cost_logger import OpenAICostLoggerViz
 
+MAX_RESTARTS = 10 # risk of overflowing system memory otherwise
+
 
 def parse_args() -> Tuple[str, str, int, bool]:
     """Parse bash arguments
@@ -86,7 +88,10 @@ def get_starting_step(exp_name: str) -> int:
     if full_path.exists():
         files = os.listdir(full_path)
         steps = [int(os.path.splitext(filename)[0]) for filename in files]
-        current_step = max(steps)
+        if len(steps) == 0:
+          current_step = 0
+        else:
+          current_step = max(steps)
     return current_step
 
 
@@ -166,7 +171,7 @@ def save_checkpoint(rs, idx: int, th: Process) -> Tuple[str, int, int]:
 
 
 if __name__ == '__main__':
-    checkpoint_freq = 200 # 1 step = 10 sec
+    checkpoint_freq = 1000 # 1 step = 10 sec
     log_path = "cost-logs" # where the simulations' prints are stored
     idx = 0
     origin, target, tot_steps, ui, browser_path, port = parse_args()
@@ -181,7 +186,9 @@ def save_checkpoint(rs, idx: int, th: Process) -> Tuple[str, int, int]:
     print(f"(Auto-Exec): Target: {target}", flush=True)
     print(f"(Auto-Exec): Total steps: {tot_steps}", flush=True)
     print(f"(Auto-Exec): Checkpoint Freq: {checkpoint_freq}", flush=True)    
-
+
+
+    number_restarts = 0 
     while current_step < tot_steps:
         try:
             steps_to_run = curr_checkpoint - current_step
@@ -207,8 +214,13 @@ def save_checkpoint(rs, idx: int, th: Process) -> Tuple[str, int, int]:
                 origin, current_step, idx = save_checkpoint(rs, idx, th)
             else:
                 shutil.rmtree(f"../../environment/frontend_server/storage/{target}") # Remove the experiment folder if no steps were run
+            number_restarts += 1
             print(f"(Auto-Exec): Error at step {current_step}", flush=True)
             print(f"(Auto-Exec): Exception {e.args[0]}", flush=True)
+            print(f"(Auto-Exec): Restart:{number_restarts} max: {MAX_RESTARTS}", flush=True)
+            if number_restarts > MAX_RESTARTS:
+                print(f"(Auto-Exec): Too many restarts, shutting down!", flush=True)
+                sys.exit(0)
         else:
             origin, current_step, idx = save_checkpoint(rs, idx, th)
             curr_checkpoint = get_new_checkpoint(current_step, tot_steps, checkpoint_freq)
@@ -229,4 +241,4 @@ def save_checkpoint(rs, idx: int, th: Process) -> Tuple[str, int, int]:
     OpenAICostLoggerViz.print_experiment_cost(experiment=exp_name, path=log_path)
     OpenAICostLoggerViz.print_total_cost(path=log_path)
     print(f"(Auto-Exec): Execution time: {datetime.now() - start_time}")
-    sys.exit(0)
+    sys.exit(0)

reverie/backend_server/persona/cognitive_modules/converse.py

@@ -208,19 +208,9 @@ def generate_inner_thought(persona, whisper):
   inner_thought = run_gpt_prompt_generate_whisper_inner_thought(persona, whisper)[0]
   return inner_thought
 
-def generate_action_event_triple(act_desp, persona): 
-  """TODO 
-
-  INPUT: 
-    act_desp: the description of the action (e.g., "sleeping")
-    persona: The Persona class instance
-  OUTPUT: 
-    a string of emoji that translates action description.
-  EXAMPLE OUTPUT: 
-    "🧈🍞"
-  """
-  if debug: print ("GNS FUNCTION: <generate_action_event_triple>")
-  return run_gpt_prompt_event_triple(act_desp, persona)[0]
+def generate_thought_triple(act_desp, persona): 
+  if debug: print ("GNS FUNCTION: <generate_thought_triple>")
+  return run_gpt_prompt_thought_triple(act_desp, persona)[0]
 
 
 def generate_poig_score(persona, event_type, description): 
@@ -245,7 +235,7 @@ def load_history_via_whisper(personas, whispers):
 
     created = persona.scratch.curr_time
     expiration = persona.scratch.curr_time + datetime.timedelta(days=30)
-    s, p, o = generate_action_event_triple(thought, persona)
+    s, p, o = generate_thought_triple(thought, persona)
     keywords = set([s, p, o])
     thought_poignancy = generate_poig_score(persona, "event", whisper)
     thought_embedding_pair = (thought, get_embedding(thought))
@@ -289,7 +279,7 @@ def open_convo_session(persona, convo_mode, safe_mode=True, direct=False, questi
 
     created = persona.scratch.curr_time
     expiration = persona.scratch.curr_time + datetime.timedelta(days=30)
-    s, p, o = generate_action_event_triple(thought, persona)
+    s, p, o = generate_thought_triple(thought, persona)
     keywords = set([s, p, o])
     thought_poignancy = generate_poig_score(persona, "event", whisper)
     thought_embedding_pair = (thought, get_embedding(thought))

reverie/backend_server/persona/cognitive_modules/execute.py

@@ -42,7 +42,7 @@ def execute(persona, maze, personas, plan):
     # to execute the current action. The goal is to pick one of them.
     target_tiles = None
 
-    print (plan)
+    print ("PLAN ----", plan)
 
     if "<persona>" in plan: 
       # Executing persona-persona interaction.
@@ -88,7 +88,8 @@ def execute(persona, maze, personas, plan):
       # string form. <maze.address_tiles> takes this and returns candidate 
       # coordinates. 
       if plan not in maze.address_tiles: 
-        maze.address_tiles["Johnson Park:park:park garden"] #ERRORRRRRRR
+        #maze.address_tiles["Johnson Park:park:park garden"] #ERRORRRRRRR
+        raise Exception("{} not in {}".format(plan, list(maze.address_tiles)))
       else: 
         target_tiles = maze.address_tiles[plan]