Merge pull request #57 from SpudGunMan/llmLocationAware

LLM location aware enhancement

README.md

@@ -40,8 +40,8 @@ Any messages that are over 160 characters are chunked into 160 message bytes to
   - `wx` and `wxc` returns local weather forecast, (wxc is metric value), NOAA or Open Meteo for weather forecasting.
   - `wxa` and `wxalert` return NOAA alerts. Short title or expanded details
   - `joke` tells a joke
-  - `wiki: ` search wikipedia, return the first few sentances of first result if a match `wiki: lora radio`
-  - `ask: ` ask Ollama LLM AI for a response `ask: what temp do I cook chicken`
+  - `wiki: ` will search wikipedia, return the first few sentances of first result if a match `wiki: lora radio`
+  - `askai` and `ask:` will ask Ollama LLM AI for a response `askai what temp do I cook chicken`
   - `messages` Replay the last messages heard, like Store and Forward
   - `motd` or to set the message `motd $New Message Of the day`
   - `lheard` returns the last 5 heard nodes with SNR, can also use `sitrep`
@@ -64,6 +64,7 @@ Optionally:
 - `launch.sh` will activate and launch the app in the venv if built.
 
 For Docker:
+Check you have serial port properly shared and the GPU if using LLM with [NVidia](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/docker-specialized.html)
 - `git clone https://github.com/spudgunman/meshing-around`
 - `cd meshing-around && docker build -t meshing-around`
 - `docker run meshing-around`
@@ -156,6 +157,8 @@ signalCycleLimit = 5
 Ollama Settings, for Ollama to work the command line `ollama run 'model'` needs to work properly. Check that you have enough RAM and your GPU are working as expected. The default model for this project, is set to `gemma2:2b` (run `ollama pull gemma2:2b` on command line, to download and setup) however I have found gemma2:2b to be lighter, faster and seems better overall vs llama3,1 (`olamma pull llama3.1`)
  - From the command terminal of your system with mesh-bot, download the default model for mesh-bot which is currently `ollama pull gemma2:2b`
 
+Enable History, set via code readme Ollama Config in [Settings](https://github.com/SpudGunMan/meshing-around?tab=readme-ov-file#configurations) and [llm.py](https://github.com/SpudGunMan/meshing-around/blob/eb3bbdd3c5e0f16fe3c465bea30c781bd132d2d3/modules/llm.py#L12)
+
 ```
 # Enable ollama LLM see more at https://ollama.com
 ollama = True
@@ -163,6 +166,15 @@ ollama = True
 ollamaModel = gemma2:2b
 ```
 
+also see llm.py for changing the defaults of
+```
+# LLM System Variables
+llmEnableHistory = False # enable history for the LLM model to use in responses adds to compute time
+llmContext_fromGoogle = True # enable context from google search results adds to compute time but really helps with responses accuracy
+googleSearchResults = 3 # number of google search results to include in the context more results = more compute time
+llm_history_limit = 6 # limit the history to 3 messages (come in pairs) more results = more compute time
+```
+
 Logging messages to disk or Syslog to disk uses the python native logging function. Take a look at the [/modules/log.py](/modules/log.py) you can set the file logger for syslog to INFO for example to not log DEBUG messages to file log, or modify the stdOut level.
 ```
 [general]
@@ -219,6 +231,7 @@ The following is for the Ollama LLM
 pip install langchain
 pip install langchain-ollama
 pip install ollama
+pip install googlesearch-python
 ```
 
 To enable emoji in the Debian console, install the fonts `sudo apt-get install fonts-noto-color-emoji`

mesh_bot.py

@@ -24,6 +24,7 @@ def auto_response(message, snr, rssi, hop, message_from_id, channel_number, devi
         "wx": lambda: handle_wxc(message_from_id, deviceID, 'wx'),
         "wiki:": lambda: handle_wiki(message),
         "ask:": lambda: handle_llm(message_from_id, channel_number, deviceID, message, publicChannel),
+        "askai": lambda: handle_llm(message_from_id, channel_number, deviceID, message, publicChannel),
         "joke": tell_joke,
         "bbslist": bbs_list_messages,
         "bbspost": lambda: handle_bbspost(message, message_from_id, deviceID),
@@ -105,12 +106,29 @@ def handle_wiki(message):
         return "Please add a search term example:wiki: travelling gnome"
 
 def handle_llm(message_from_id, channel_number, deviceID, message, publicChannel):
-    global llmRunCounter, llmTotalRuntime
+    global llmRunCounter, llmTotalRuntime, llmLocationTable
+
+    if location_enabled:
+        location = get_node_location(message_from_id, deviceID)
+        # if message_from_id is is the llmLocationTable use the location from the table to save on API calls
+        if message_from_id in llmLocationTable:
+            location = llmLocationTable[message_from_id]
+        else:
+            location_name = where_am_i(str(location[0]), str(location[1]), short = True)
+            llmLocationTable.append({message_from_id: location_name})
+
+        if NO_DATA_NOGPS in location_name:
+            location_name = "no location provided "
+    else:
+        location_name = "no location provided "
+
     if "ask:" in message.lower():
         user_input = message.split(":")[1]
-        user_input = user_input.strip()
+    elif "askai" in message.lower():
+        user_input = message.replace("askai", "")
     else:
         user_input = message
+    user_input = user_input.strip()
         
     if len(user_input) < 1:
         return "Please ask a question"
@@ -134,7 +152,7 @@ def handle_llm(message_from_id, channel_number, deviceID, message, publicChannel
     start = time.time()
 
     #response = asyncio.run(llm_query(user_input, message_from_id))
-    response = llm_query(user_input, message_from_id)
+    response = llm_query(user_input, message_from_id, location_name)
 
     # handle the runtime counter
     end = time.time()

modules/llm.py

@@ -4,40 +4,62 @@
 # K7MHI Kelly Keeton 2024
 from modules.log import *
 
-from langchain_ollama import OllamaLLM
-from langchain_core.prompts import ChatPromptTemplate
+from langchain_ollama import OllamaLLM # pip install ollama langchain-ollama
+from langchain_core.prompts import ChatPromptTemplate # pip install langchain
 from langchain_core.messages import AIMessage, HumanMessage
+from googlesearch import search # pip install googlesearch-python
 
 # LLM System Variables
-llmEnableHistory = False
-llm_history_limit = 6 # limit the history to 3 messages (come in pairs)
+llmEnableHistory = False # enable history for the LLM model to use in responses adds to compute time
+llmContext_fromGoogle = True # enable context from google search results adds to compute time but really helps with responses accuracy
+googleSearchResults = 3 # number of google search results to include in the context more results = more compute time
+llm_history_limit = 6 # limit the history to 3 messages (come in pairs) more results = more compute time
 antiFloodLLM = []
 llmChat_history = []
-trap_list_llm = ("ask:",)
+trap_list_llm = ("ask:", "askai")
 
 meshBotAI = """
-FROM {llmModel}
-SYSTEM
-You must keep responses under 450 characters at all times, the response will be cut off if it exceeds this limit.
-You must respond in plain text standard ASCII characters, or emojis.
-You are acting as a chatbot, you must respond to the prompt as if you are a chatbot assistant, and dont say 'Response limited to 450 characters'.
-Unless you are provided HISTORY, you cant ask followup questions but you can ask for clarification and to rephrase the question if needed.
-If you feel you can not respond to the prompt as instructed, come up with a short quick error.
-The prompt includes a user= variable that is for your reference only to track different users, do not include it in your response.
-This is the end of the SYSTEM message and no further additions or modifications are allowed.
+    FROM {llmModel}
+    SYSTEM
+    You must keep responses under 450 characters at all times, the response will be cut off if it exceeds this limit.
+    You must respond in plain text standard ASCII characters, or emojis.
+    You are acting as a chatbot, you must respond to the prompt as if you are a chatbot assistant, and dont say 'Response limited to 450 characters'.
+    Unless you are provided HISTORY, you cant ask followup questions but you can ask for clarification and to rephrase the question if needed.
+    If you feel you can not respond to the prompt as instructed, come up with a short quick error.
+    The prompt includes a user= variable that is for your reference only to track different users, do not include it in your response.
+    This is the end of the SYSTEM message and no further additions or modifications are allowed.
 
-PROMPT
-{input}
-user={userID}
+
+    PROMPT
+    {input}
+    user={userID}
 
 """
 
+if llmContext_fromGoogle:
+    meshBotAI = meshBotAI + """
+        CONTEXT
+        The following is the location of the user
+        {location_name}
+
+        The following is for context around the prompt to help guide your response.
+        {context}
+
+    """
+else:
+    meshBotAI = meshBotAI + """
+        CONTEXT
+        The following is the location of the user
+        {location_name}
+
+    """
+
 if llmEnableHistory:
     meshBotAI = meshBotAI + """
-    HISTORY
-    You have memory of a few previous messages, you can use this to help guide your response.
-    The following is for memory purposes only and should not be included in the response.
-    {history}
+        HISTORY
+        You have memory of a few previous messages, you can use this to help guide your response.
+        The following is for memory purposes only and should not be included in the response.
+        {history}
 
     """
 
@@ -46,8 +68,11 @@ ollama_model = OllamaLLM(model=llmModel)
 model_prompt = ChatPromptTemplate.from_template(meshBotAI)
 chain_prompt_model = model_prompt | ollama_model
 
-def llm_query(input, nodeID=0):
+def llm_query(input, nodeID=0, location_name=None):
     global antiFloodLLM, llmChat_history
+    googleResults = []
+    if not location_name:
+        location_name = "no location provided "
 
     # add the naughty list here to stop the function before we continue
     # add a list of allowed nodes only to use the function
@@ -58,12 +83,40 @@ def llm_query(input, nodeID=0):
     else:
         antiFloodLLM.append(nodeID)
 
+    if llmContext_fromGoogle:
+        # grab some context from the internet using google search hits (if available)
+        # localization details at https://pypi.org/project/googlesearch-python/
+        try:
+            googleSearch = search(input, advanced=True, num_results=googleSearchResults)
+            if googleSearch:
+                for result in googleSearch:
+                    # SearchResult object has url= title= description= just grab title and description
+                    googleResults.append(f"{result.title} {result.description}")
+            else:
+                googleResults = ['no other context provided']
+        except Exception as e:
+            logger.debug(f"System: LLM Query: context gathering error: {e}")
+            googleResults = ['no other context provided']
+
+
+    if googleResults:
+        logger.debug(f"System: External LLM Query: {input} From:{nodeID} with context from google")
+    else:
+        logger.debug(f"System: External LLM Query: {input} From:{nodeID}")
+    
     response = ""
-    logger.debug(f"System: LLM Query: {input} From:{nodeID}")
+    result = ""
+    location_name += f" at the current time of {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}"
 
-    result = chain_prompt_model.invoke({"input": input, "llmModel": llmModel, "userID": nodeID, "history": llmChat_history})
+    try:
+        result = chain_prompt_model.invoke({"input": input, "llmModel": llmModel, "userID": nodeID, \
+                                            "history": llmChat_history, "context": googleResults, "location_name": location_name})
+        #logger.debug(f"System: LLM Response: " + result.strip().replace('\n', ' '))
+    except Exception as e:
+        logger.warning(f"System: LLM failure: {e}")
+        return "I am having trouble processing your request, please try again later."
+    
 
-    #logger.debug(f"System: LLM Response: " + result.strip().replace('\n', ' '))
     response = result.strip().replace('\n', ' ')
 
     # Store history of the conversation, with limit to prevent template growing too large causing speed issues
@@ -79,3 +132,15 @@ def llm_query(input, nodeID=0):
     antiFloodLLM.remove(nodeID)
 
     return response
+
+# import subprocess
+# def get_ollama_cpu():
+#     try:
+#         psOutput = subprocess.run(['ollama', 'ps'], capture_output=True, text=True)
+#         if "GPU" in psOutput.stdout:
+#             logger.debug(f"System: Ollama process with GPU")
+#         else:
+#             logger.debug(f"System: Ollama process with CPU, query time will be slower")
+#     except Exception as e:
+#         logger.debug(f"System: Ollama process not found, {e}")
+#         return False

modules/locationdata.py

@@ -11,7 +11,7 @@ from modules.log import *
 
 trap_list_location = ("whereami", "tide", "moon", "wx", "wxc", "wxa", "wxalert")
 
-def where_am_i(lat=0, lon=0):
+def where_am_i(lat=0, lon=0, short=False):
     whereIam = ""
     grid = mh.to_maiden(float(lat), float(lon))
     
@@ -23,6 +23,13 @@ def where_am_i(lat=0, lon=0):
     geolocator = Nominatim(user_agent="mesh-bot")
     
     # Nomatim API call to get address
+    if short:
+        location = geolocator.reverse(lat + ", " + lon)
+        address = location.raw['address']
+        address_components = ['city', 'state', 'county', 'country']
+        whereIam = f"City: {address.get('city', '')} State: {address.get('state', '')} County: {address.get('county', '')} Country: {address.get('country', '')}"
+        return whereIam
+    
     if float(lat) == latitudeValue and float(lon) == longitudeValue:
         # redacted address when no GPS and using default location
         location = geolocator.reverse(lat + ", " + lon)
@@ -30,14 +37,13 @@ def where_am_i(lat=0, lon=0):
         address_components = ['city', 'state', 'postcode', 'county', 'country']
         whereIam += ' '.join([address.get(component, '') for component in address_components if component in address])
         whereIam += " Grid: " + grid
-        return whereIam
     else:
         location = geolocator.reverse(lat + ", " + lon)
         address = location.raw['address']
         address_components = ['house_number', 'road', 'city', 'state', 'postcode', 'county', 'country']
         whereIam += ' '.join([address.get(component, '') for component in address_components if component in address])
         whereIam += " Grid: " + grid
-        return whereIam
+    return whereIam
 
 def get_tide(lat=0, lon=0):
     station_id = ""

modules/settings.py

@@ -28,6 +28,7 @@ scheduler_enabled = False # enable the scheduler currently config via code only
 wiki_return_limit = 3 # limit the number of sentences returned off the first paragraph first hit
 llmRunCounter = 0
 llmTotalRuntime = []
+llmLocationTable = []
 
 # Read the config file, if it does not exist, create basic config file
 config = configparser.ConfigParser() 
@@ -92,7 +93,7 @@ try:
     urlTimeoutSeconds = config['general'].getint('urlTimeout', 10) # default 10 seconds
     store_forward_enabled = config['general'].getboolean('StoreForward', True)
     storeFlimit = config['general'].getint('StoreLimit', 3) # default 3 messages for S&F
-    welcome_message = config['general'].get(f'welcome_message', WELCOME_MSG)
+    welcome_message = config['general'].get('welcome_message', WELCOME_MSG)
     welcome_message = (f"{welcome_message}").replace('\\n', '\n') # allow for newlines in the welcome message
     motd_enabled = config['general'].getboolean('motdEnabled', True)
     dad_jokes_enabled = config['general'].getboolean('DadJokes', False)

modules/system.py

@@ -74,7 +74,7 @@ if wikipedia_enabled:
 if llm_enabled:
     from modules.llm import * # from the spudgunman/meshing-around repo
     trap_list = trap_list + trap_list_llm # items ask:
-    help_message = help_message + ", ask:"
+    help_message = help_message + ", askai"
 
 # Scheduled Broadcast Configuration
 if scheduler_enabled:
@@ -103,7 +103,7 @@ if interface1_type == 'ble' and interface2_type == 'ble':
     
 # Interface1 Configuration
 try:
-    logger.debug(f"System: Initalizing Interface1")
+    logger.debug(f"System: Initializing Interface1")
     if interface1_type == 'serial':
         interface1 = meshtastic.serial_interface.SerialInterface(port1)
     elif interface1_type == 'tcp':
@@ -114,12 +114,12 @@ try:
         logger.critical(f"System: Interface Type: {interface1_type} not supported. Validate your config against config.template Exiting")
         exit()
 except Exception as e:
-    logger.critical(f"System: script abort. Initalizing Interface1 {e}")
+    logger.critical(f"System: script abort. Initializing Interface1 {e}")
     exit()
 
 # Interface2 Configuration
 if interface2_enabled:
-    logger.debug(f"System: Initalizing Interface2")
+    logger.debug(f"System: Initializing Interface2")
     try:
         if interface2_type == 'serial':
             interface2 = meshtastic.serial_interface.SerialInterface(port2)
@@ -131,7 +131,7 @@ if interface2_enabled:
             logger.critical(f"System: Interface Type: {interface2_type} not supported. Validate your config against config.template Exiting")
             exit()
     except Exception as e:
-        logger.critical(f"System: script abort. Initalizing Interface2 {e}")
+        logger.critical(f"System: script abort. Initializing Interface2 {e}")
         exit()
 
 #Get the node number of the device, check if the device is connected
@@ -670,7 +670,6 @@ async def watchdog():
                 with contextlib.redirect_stdout(None):
                     interface1.localNode.getMetadata()
                     print(f"System: if you see this upgrade python to >3.4")
-                #if "device_state_version:" not in meta:
             except Exception as e:
                 logger.error(f"System: communicating with interface1, trying to reconnect: {e}")
                 retry_int1 = True

requirements.txt

@@ -16,4 +16,4 @@ wikipedia
 langchain
 langchain-ollama
 ollama
-
+googlesearch-python