Make microchain agent runnable with local models via Ollama (#98)

* Add notes on different model performances of the microchain agent
* Merge main

prediction_market_agent/agents/microchain_agent/microchain_agent.py

@@ -1,3 +1,4 @@
+import typer
 from functions import MARKET_FUNCTIONS, MISC_FUNCTIONS
 from microchain import LLM, Agent, Engine, OpenAIChatGenerator
 from microchain.functions import Reasoning, Stop
@@ -8,32 +9,41 @@
 )
 from prediction_market_agent.utils import APIKeys
 
-engine = Engine()
-engine.register(Reasoning())
-engine.register(Stop())
-for function in MISC_FUNCTIONS:
-    engine.register(function())
-for function in MARKET_FUNCTIONS:
-    engine.register(function(market_type=MarketType.OMEN))
-for function in OMEN_FUNCTIONS:
-    engine.register(function())
 
-generator = OpenAIChatGenerator(
-    model="gpt-4-turbo-preview",
-    api_key=APIKeys().openai_api_key.get_secret_value(),
-    api_base="https://api.openai.com/v1",
-    temperature=0.7,
-)
-agent = Agent(llm=LLM(generator=generator), engine=engine)
-agent.prompt = f"""Act as a agent to maximise your profit. You can use the following functions:
- 
-{engine.help}
- 
- 
-Only output valid Python function calls.
- 
-"""
+def main(
+    api_base: str = "https://api.openai.com/v1",
+    model: str = "gpt-4-turbo-preview",
+) -> None:
+    engine = Engine()
+    engine.register(Reasoning())
+    engine.register(Stop())
+    for function in MISC_FUNCTIONS:
+        engine.register(function())
+    for function in MARKET_FUNCTIONS:
+        engine.register(function(market_type=MarketType.OMEN))
+    for function in OMEN_FUNCTIONS:
+        engine.register(function())
+
+    generator = OpenAIChatGenerator(
+        model=model,
+        api_key=APIKeys().openai_api_key.get_secret_value(),
+        api_base=api_base,
+        temperature=0.7,
+    )
+    agent = Agent(llm=LLM(generator=generator), engine=engine)
+    agent.prompt = f"""Act as a agent to maximise your profit. You can use the following functions:
+    
+    {engine.help}
+    
+    
+    Only output valid Python function calls.
+    
+    """
+
+    agent.bootstrap = ['Reasoning("I need to reason step-by-step")']
+    agent.run(iterations=10)
+    # generator.print_usage() # Waiting for microchain release
+
 
-agent.bootstrap = ['Reasoning("I need to reason step-by-step")']
-agent.run(iterations=10)
-generator.print_usage()
+if __name__ == "__main__":
+    typer.run(main)

prediction_market_agent/agents/microchain_agent/model_notes.md

@@ -0,0 +1,52 @@
+# Microchain Agent Model Behaviour Diary
+
+## Proprietary models
+
+### GPT4
+
+- Makes many reasoning steps, with coherent and good reasoning w.r.t betting strategy
+- Almost always gets function calls correct
+- Seems keen on betting large amounts, even when instructed not to!
+- Seems keen to `Stop` program after max a couple bets. Doesn't use some of the functions (selling, getting existing positions)
+
+## Local models
+
+### Setup
+
+- Instructions are for Ollama, but you can use any library that allows you to set up a local OpenAI-compatible server.
+- Download Ollama [here](https://ollama.com/download/mac)
+- In another terminal, run `ollama serve` to start the server. You can set the address and port with the `OLLAMA_HOST` env var, e.g.:
+
+```bash
+OLLAMA_HOST=127.0.0.1:11435 ollama serve`
+```
+
+- Run the script, passing in the API address and model name as arguments. Note that you must have downloaded these model weights in advance via `ollama run <model_name`:
+
+```bash
+python prediction_market_agent/agents/microchain_agent/microchain_agent.py --api-base "http://localhost:11435/v1" --model "nous-hermes2:latest"
+```
+
+Note that the first call to the model will be slow, as the model weights are loaded from disk into GPU memory/RAM.
+
+## [mixtral:8x7b-instruct-v0.1-q3_K_S](https://ollama.com/library/mixtral:8x7b-instruct-v0.1-q3_K_S)
+
+- Promising! Outputs some coherent reasoning. Chains several function calls together before starting to lose its way.
+- Made several bad function calls, but followed up with reasoning to fix the function call, then made correct one.
+- Didn't have a good view of all functions available. e.g.
+  - tried to use `EndChat()` instead of `Stop()`
+  - iterated through all markets to get market.p_yes, but didn't try to call a mech to predict its own p_yes.
+- Questionable reasoning w.r.t betting strategy. Stated that a market with a large |p_yes - 0.5| was more likely to be mis-priced.
+
+## [llama3:latest](https://ollama.com/library/llama3)
+
+- Couldn't get any useful function calls from it.
+- Often replied with an empty string, aborting the program
+- Couldn't recover after an incorrect function call
+
+## [nous-hermes2:latest](https://ollama.com/library/nous-hermes2)
+
+- Took some system prompt massaging for it to get going
+- Was able to recover from some bad function calls
+- Correctly called a mech to predict its own p_yes, but reasoning about what to do with result unraveled
+- Always falls over before placing a bet

prediction_market_agent/agents/microchain_agent/omen_functions.py

@@ -3,7 +3,7 @@
     redeem_from_all_user_positions,
 )
 
-from prediction_market_agent.agents.microchain_agent.utils import MicrochainAPIKeys
+from prediction_market_agent.utils import APIKeys
 
 
 class RedeemWinningBets(Function):
@@ -16,7 +16,7 @@ def example_args(self) -> list[str]:
         return []
 
     def __call__(self) -> None:
-        redeem_from_all_user_positions(MicrochainAPIKeys().bet_from_private_key)
+        redeem_from_all_user_positions(APIKeys().bet_from_private_key)
 
 
 # Functions that interact exclusively with Omen prediction markets