feat: re-enable Exa search

docs: add information regarding what models to use in `.env.example`

From my testing, Mistral was the best at telling me the current Android + iOS versions. It actually used DuckDuckGo. However, it was still slightly off for Android. Might have been better to only provide Exa instead. In the future, it may be beneficial to add an argument that disables DuckDuckGo or just disable DuckDuckGo if Exa is being used.

.env.example

@@ -28,8 +28,10 @@ LLM_BASE_URL="https://openrouter.ai/api/v1"
 # The model to use
 # For openrouter.ai, you can check the available models at https://openrouter.ai/docs#models
 # For Ollama, you can check the available models at https://ollama.com/library
+# So far, for the function-calling models I've tested, `mistral:7b-instruct-v0.3-q6_K` works the best (show_tool_calls disabled)
+# I use this as a reference for the quantization methods: https://github.com/ggerganov/llama.cpp/discussions/2094#discussioncomment-6351796
 LLM_MODEL = "mistralai/mistral-7b-instruct:free"
-
+SHOW_TOOL_CALLS = false
 # Settings for Gmail
 IMAP_HOST=imap.gmail.com
 IMAP_PORT=993

llmail/utils/cli_args.py

@@ -38,26 +38,6 @@ def set_argparse():
     # Subcommand: list-folders
     _ = subparsers.add_parser("list-folders", help="List all folders in the IMAP account and exit")
     # General arguments
-    argparser.add_argument(
-        "--log-level",
-        "-l",
-        help="Log level",
-        default=os.getenv("LOG_LEVEL") if os.getenv("LOG_LEVEL") else "INFO",
-    )
-    argparser.add_argument(
-        "--redact-email-addresses",
-        help="Replace email addresses with '[redacted]' in logs",
-        action="store_true",
-        default=(
-            True
-            if (
-                os.getenv("REDACT_EMAIL_ADDRESSES")
-                and os.getenv("REDACT_EMAIL_ADDRESSES").lower() == "true"
-                and os.getenv("REDACT_EMAIL_ADDRESSES").lower() != "false"
-            )
-            else False
-        ),
-    )
     argparser.add_argument(
         "--exa-api-key",
         help="Exa API key",
@@ -168,6 +148,42 @@ def set_argparse():
         default=(os.getenv("MESSAGE_ID_DOMAIN") if os.getenv("MESSAGE_ID_DOMAIN") else None),
     )
 
+    # Debuggiongm arguments
+    debug = argparser.add_argument_group("Debugging")
+    debug.add_argument(
+        "--log-level",
+        "-l",
+        help="Log level",
+        default=os.getenv("LOG_LEVEL") if os.getenv("LOG_LEVEL") else "INFO",
+    )
+    debug.add_argument(
+        "--redact-email-addresses",
+        help="Replace email addresses with '[redacted]' in logs",
+        action="store_true",
+        default=(
+            True
+            if (
+                os.getenv("REDACT_EMAIL_ADDRESSES")
+                and os.getenv("REDACT_EMAIL_ADDRESSES").lower() == "true"
+                and os.getenv("REDACT_EMAIL_ADDRESSES").lower() != "false"
+            )
+            else False
+        ),
+    )
+    debug.add_argument(
+        "--show-tool-calls",
+        help="Pass show_tool_calls=True to phidata",
+        action="store_true",
+        default=(
+            True
+            if (
+                os.getenv("SHOW_TOOL_CALLS")
+                and os.getenv("SHOW_TOOL_CALLS").lower() == "true"
+                and os.getenv("SHOW_TOOL_CALLS").lower() != "false"
+            )
+            else False
+        ),
+    )
     check_required_args(
         [
             "imap_host",

llmail/utils/responding.py

@@ -12,7 +12,7 @@
 from phi.llm.ollama import Ollama
 from phi.tools.exa import ExaTools
 from phi.tools.website import WebsiteTools
-
+import ollama
 
 # Uses utils/__init__.py to import from utils/logging.py and utils/cli_args.py respectively
 from llmail.utils import logger, args, bot_email
@@ -166,8 +166,8 @@ def fetch_and_process_emails(
                 WebsiteTools(),
                 DuckDuckGo(search=True, news=True)
             ]
-            # if args.exa_api_key is not None:
-            #     tools.append(ExaTools(api_key=args.exa_api_key, use_autoprompt=True, highlights=True))
+            if args.exa_api_key is not None:
+                tools.append(ExaTools(api_key=args.exa_api_key, highlights=True))
             if args.no_tools:
                 tools = []
             # Chose how to send the request to the provider
@@ -179,20 +179,31 @@ def fetch_and_process_emails(
                         base_url=args.llm_base_url,
                     )
                 case "ollama":
+                    ollama_client = ollama.Client(host=args.llm_base_url)
+                    for model in ollama_client.list()["models"]:
+                        if model["name"] == args.llm_model:
+                            logger.debug(f"{args.llm_model} is already downloaded")
+                            break
+                    else:
+                        logger.info(f"Downloading {args.llm_model}")
+                        ollama_client.pull(model=args.llm_model)
                     llm = Ollama(
                         model=args.llm_model,
                         host=args.llm_base_url,
                     )
             assistant = Assistant(
                 llm=llm,
                 tools=tools,
-                show_tool_calls=False,
+                show_tool_calls=args.show_tool_calls,
                 # additional_messages=[
                 #     {
                 #         "role": "system",
                 #         "content": "Functions generally need arguments. Only provide your final iteration to the user. For example, if you're searching for a website, don't tell the user that you're searching for the website. Just provide the final result. Remember, the user is only going to see the final response, not the steps you took to get there. Only provide the final result. Try to use the most appropriate tool for the task. For example, if the user asks about a website, try searching for it and scraping it too. For URLs, ensure the protocol is included (e.g. https://). Use the functions to ensure that the information is accurate and up-to-date",
                 #     },
                 # ],
+                tool_call_limit=10,
+                debug_mode=False,
+                prevent_hallucinations=True,
             )
             send_reply(
                 thread=tracking.get_thread_history(client, email_thread),