Merge pull request #43 from ssoudan/upd

GPT 3.5 Turbo 0613
ssoudan · Jun 14, 2023 · c8d2e5a · c8d2e5a
2 parents e0ca15e + 2fb4b4e
commit c8d2e5a
Show file tree

Hide file tree

Showing 30 changed files with 479 additions and 360 deletions.
diff --git a/Cargo.lock b/Cargo.lock
diff --git a/README.md b/README.md
@@ -146,7 +146,7 @@ parameters:
 ```
 
 =============
-# Action SandboxedPython result:
+# Action SandboxedPython response:
 ```yaml
 stdout: |
   The solution to x^2 - 5x + 6 = 0 is [2, 3]
@@ -235,7 +235,7 @@ parameters:
 ```
 
 =============   2 messages in the chat history =============
-# Action result:
+# Action response:
 ```yaml
 stdout: |
   The temperature in San Jose, California, USA is 12.78°C and 10 Day Weather-San Jose, CA ; Mon 17 · 59°. 1%. NW 13 mph. Cloudy. High 59F. Winds NW at 10 to 15 mph. ; Tue 18 · 61°. 7%. NW 14 mph. Partly cloudy. High 61F. Winds ...
@@ -373,7 +373,7 @@ parameters:
 
 WORLD:
 <pre><code>
-# Action result:
+# Action response:
 ```yaml
 status: 0
 stdout: |
@@ -432,7 +432,7 @@ parameters:
 
 WORLD:
 <pre><code> 
-# Action result: 
+# Action response: 
 ```yaml
 rooms: []
 ```
@@ -462,7 +462,7 @@ parameters:
 
 WORLD:
 <pre><code> 
-# Action result: 
+# Action response: 
 ```yaml
 rooms:
 - name: Living
@@ -530,7 +530,7 @@ parameters:
 
 WORLD:
 <pre><code> 
-# Action result: 
+# Action response: 
 ```yaml
 lights:
 - id: '23'
@@ -781,7 +781,7 @@ parameters:
 
 WORLD:
 <pre><code> 
-# Action result: 
+# Action response: 
 ```yaml
 rooms:
 - name: Kitchen
@@ -847,7 +847,7 @@ input:
 
 WORLD:
 <pre><code> 
-# Action result: 
+# Action response: 
 ```yaml
 lights:
 - id: '26'

diff --git a/experiments/analysis.ipynb b/experiments/analysis.ipynb
@@ -439,9 +439,11 @@
    "metadata": {},
    "outputs": [],
    "source": [
+    "weak_prior = BetaPrior(alpha=100, beta=100)\n",
+    "strong_prior = BetaPrior(alpha=10000, beta=10000)\n",
     "trace_weak, trace_strong = run_scenario_twovariant(data, \n",
-    "                                                   weak_prior=BetaPrior(alpha=100, beta=100), \n",
-    "                                                   strong_prior=BetaPrior(alpha=10000, beta=10000))"
+    "                                                   weak_prior=weak_prior,\n",
+    "                                                   strong_prior=strong_prior)"
    ]
   },
   {
@@ -451,13 +453,6 @@
    "source": [
     "$\\textrm{relative\\_uplift} = (\\textrm{treatment} - \\textrm{control}) / \\textrm{control}$"
    ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": []
   }
  ],
  "metadata": {

diff --git a/sapiens/Cargo.toml b/sapiens/Cargo.toml
@@ -29,7 +29,7 @@ serde_yaml = "0.9.21"
 
 regex = "1.8.4"
 
-clap = { version = "4.3.2", optional = true }
+clap = { version = "4.3.4", optional = true }
 
 # OpenAI API - OpenAI and lm-sys/FastChat
 tiktoken-rs = { version = "0.4.2", features = ["async-openai"] }

diff --git a/sapiens/src/chains/agents/ooda/multistep.rs b/sapiens/src/chains/agents/ooda/multistep.rs
@@ -12,8 +12,9 @@ use crate::{chains, prompt, SapiensConfig, WeakRuntimeObserver};
 const PREFIX: &str = r"You are part of a group of cooperating assistants named Sapiens. Use available tools to answer the question as best as you can.
 You will collectively proceed iteratively using an OODA loop. Don't overstep your role.
 
-- Action result will be provided. 
-- Never produce the result of an Action. 
+- Action response will be provided. 
+- Never produce the response of an Action. 
+- Only use YAML for the Action.
 - The loop will repeated until you have the answer to the original question. 
 - No task is complete until the Conclude Tool is used to provide the answer. 
 ";
@@ -68,7 +69,7 @@ tool_name: <ToolName>
 parameters:
     <...>  
 ```
-We will take further action based on the result.
+We will take further action based on the response.
 ====================
 
 Notes: 
@@ -87,7 +88,7 @@ const OBSERVER_PROTO_SECOND_INPUT: &str = r#"
 ## Orientation:
 - SandboxedPython can be used to sort the list.
 - I need to provide only the `tool_name` and `parameters` fields for the SandboxedPython Tool.
-- I expect the result of the Action to contains the field `stdout` with the sorted list and `stderr` empty.
+- I expect the response of the Action to contains the field `stdout` with the sorted list and `stderr` empty.
 - I need to use the Conclude Tool to terminate the task when I have the sorted list in plain text.
 ## Decision:
 - We can use the sorted() function of Python to sort the list.
@@ -100,8 +101,8 @@ parameters:
     sorted_list = sorted(lst)
     print(f"The sorted list is {sorted_list}")
 ```
-We will take further action based on the result.
-# Action SandboxedPython result:
+We will take further action based on the response.
+# Action SandboxedPython response:
 ```yaml
 stdout: |
   The sorted list is [1, 2, 3, 4, 5]
@@ -112,15 +113,15 @@ stderr: ''
 const OBSERVER_PROTO_SECOND_RESPONSE: &str = r"
 ## Observations:
 - We needed to sort the list in ascending order.
-- We have the result of the Action.
+- We have the response of the Action.
 - We have the sorted list: [1, 2, 3, 4, 5].
 ";
 
 const ORIENTER_PROTO_INITIAL_RESPONSE: &str = r#"
 ## Orientation:
 - SandboxedPython can be used to sort the list.
 - I need to provide only the `tool_name` and `parameters` fields for the SandboxedPython Tool.
-- I expect the result of the Action to contains the field `stdout` with the sorted list and `stderr` empty.
+- I expect the response of the Action to contains the field `stdout` with the sorted list and `stderr` empty.
 - I need to use the Conclude Tool to terminate the task when I have the sorted list in plain text.
 "#;
 
@@ -136,16 +137,16 @@ parameters:
     sorted_list = sorted(lst)
     print(f"The sorted list is {sorted_list}")
 ```
-We will take further action based on the result.
-# Action SandboxedPython result:
+We will take further action based on the response.
+# Action SandboxedPython response:
 ```yaml
 stdout: |
   The sorted list is [1, 2, 3, 4, 5]
 stderr: ''
 ```
 ## Observations:
 - We needed to sort the list in ascending order.
-- We have the result of the Action.
+- We have the response of the Action.
 - We have the sorted list: [1, 2, 3, 4, 5].
 "#;
 
@@ -170,16 +171,16 @@ parameters:
     sorted_list = sorted(lst)
     print(f"The sorted list is {sorted_list}")
 ```
-We will take further action based on the result.
-# Action SandboxedPython result:
+We will take further action based on the response.
+# Action SandboxedPython response:
 ```yaml
 stdout: |
   The sorted list is [1, 2, 3, 4, 5]
 stderr: ''
 ```
 ## Observations:
 - We needed to sort the list in ascending order.
-- We have the result of the Action.
+- We have the response of the Action.
 - We have the sorted list: [1, 2, 3, 4, 5].
 ## Orientation:
 - I know the answer to the original question.
@@ -201,19 +202,19 @@ parameters:
     sorted_list = sorted(lst)
     print(f"The sorted list is {sorted_list}")
 ```
-That's it for now. We will take further action based on the result.
+That's it for now. We will take further action based on the response.
 "#;
 
 const ACTOR_PROTO_SECOND_INPUT: &str = r#"
-# Action SandboxedPython result:
+# Action SandboxedPython response:
 ```yaml
 stdout: |
   The sorted list is [1, 2, 3, 4, 5]
 stderr: ''
 ```
 ## Observations:
 - We needed to sort the list in ascending order.
-- We have the result of the Action.
+- We have the response of the Action.
 - We have the sorted list: [1, 2, 3, 4, 5].
 ## Orientation:
 - I know the answer to the original question.
@@ -830,7 +831,7 @@ mod tests {
             content: indoc! {r#"
             ## Observations:
             - We needed to sort the list in ascending order.
-            - We have the result of the Action.
+            - We have the response of the Action.
             - We have the sorted list: [1, 2, 3, 4, 5].
             "#
             }
@@ -858,7 +859,7 @@ mod tests {
             content: indoc! {r#"
             ## Observations:
             - We needed to sort the list in ascending order.
-            - We have the result of the Action.
+            - We have the response of the Action.
             - We have the sorted list: [1, 2, 3, 4, 5].
             "#
             }
@@ -898,7 +899,7 @@ mod tests {
             content: indoc! {r#"
             ## Observations:
             - We needed to sort the list in ascending order.
-            - We have the result of the Action.
+            - We have the response of the Action.
             - We have the sorted list: [1, 2, 3, 4, 5].
             "#
             }
@@ -964,7 +965,7 @@ mod tests {
             ## Orientation:
             - SandboxedPython can be used to sort the list.
             - I need to provide only the `tool_name` and `parameters` fields for the SandboxedPython Tool.
-            - I expect the result of the Action to contains the field `stdout` with the sorted list and `stderr` empty.
+            - I expect the response of the Action to contains the field `stdout` with the sorted list and `stderr` empty.
             - I need to use the Conclude Tool to terminate the task when I have the sorted list in plain text."#
             }.trim().to_string(),
             usage: None,
@@ -991,7 +992,7 @@ mod tests {
                 sorted_list = sorted(lst)
                 print(f"The sorted list is {sorted_list}")
             ```
-            That's it for now. We will take further action based on the result.            
+            That's it for now. We will take further action based on the response.            
             "#
             }
             .trim()

diff --git a/sapiens/src/chains/agents/ooda/one_step.rs b/sapiens/src/chains/agents/ooda/one_step.rs
@@ -17,8 +17,9 @@ pub struct Agent {
 const PREFIX: &str = r"You are Sapiens, a large language model assisting the WORLD. Use available tools to answer the question as best as you can.
 You will proceed iteratively using an OODA loop.
 
-- Action result will be provided to you. 
-- Never produce the result of an Action. 
+- Action response will be provided to you. 
+- Never produce the response of an Action. 
+- Only use YAML for the Action.
 - The loop will repeated until you have the answer to the original question. 
 - No task is complete until the Conclude Tool is used to provide the answer.
 - You cannot use jinja2 templating in your response. Be concise. 
@@ -49,13 +50,13 @@ tool_name: <ToolName>
 parameters:
     <...>  
 ```
-We will take further action based on the result.
+We will take further action based on the response.
 ====================
 
 Notes: 
 - Action has the following fields: `tool_name` and `parameters` ONLY.
 - `parameters` uses the format specified for the Tool.
-- `result_fields` is the format you can expect of the result of the Action. You can use this to orient yourself but never use it in your response.
+- `responses_content` is the format you can expect of the response of the Action. You can use this to orient yourself but never use it in your response.
 - One Action at a time. No more. No less.
 ";
 
@@ -66,7 +67,7 @@ const PROTO_EXCHANGE_2: &str = r#"
 ## Orientation:
 - SandboxedPython can be used to sort the list.
 - I need to provide only the `tool_name` and `parameters` fields for the SandboxedPython Tool.
-- I expect the result of the Action to contains the field `stdout` with the sorted list and `stderr` empty.
+- I expect the response of the Action to contains the field `stdout` with the sorted list and `stderr` empty.
 - I need to use the Conclude Tool to terminate the task when I have the sorted list in plain text.
 ## Decision:
 - We can use the sorted() function of Python to sort the list.
@@ -79,11 +80,11 @@ parameters:
     sorted_list = sorted(lst)
     print(f"The sorted list is {sorted_list}")
 ```
-We will take further action based on the result.
+We will take further action based on the response.
 "#;
 
 const PROTO_EXCHANGE_3: &str = r"
-# Action SandboxedPython result:
+# Action SandboxedPython response:
 ```yaml
 stdout: |
   The sorted list is [1, 2, 3, 4, 5]
@@ -94,7 +95,7 @@ stderr: ''
 const PROTO_EXCHANGE_4: &str = r"
 ## Observations:
 - We needed to sort the list in ascending order.
-- We have the result of the Action.
+- We have the response of the Action.
 - We have the sorted list: [1, 2, 3, 4, 5].
 ## Orientation:
 - I know the answer to the original question.
@@ -198,7 +199,7 @@ impl Agent {
                         role: Role::User,
                     };
 
-                    // Add the result to the chat history
+                    // Add the response to the chat history
                     chat_history.add_chitchat(entry).await;
                 }
                 _ => {
@@ -295,7 +296,7 @@ mod tests {
             ## Orientation:
             - SandboxedPython can be used to sort the list.
             - I need to provide only the `tool_name` and `parameters` fields for the SandboxedPython Tool.
-            - I expect the result of the Action to contains the field `stdout` with the sorted list and `stderr` empty.
+            - I expect the response of the Action to contains the field `stdout` with the sorted list and `stderr` empty.
             - I need to use the Conclude Tool to terminate the task when I have the sorted list in plain text.
             ## Decision:
             - We can use the sorted() function of Python to sort the list.
@@ -308,7 +309,7 @@ mod tests {
                 sorted_list = sorted(lst)
                 print(f"The sorted list is {sorted_list}")
             ```
-            We will take further action based on the result.            
+            We will take further action based on the response.            
             "#
             }.to_string(),
             usage: None,