superagent-ai · elisalimli · Mar 30, 2024 · Mar 30, 2024 · Mar 30, 2024 · Mar 30, 2024
diff --git a/fern/apis/prod/openapi/openapi.yaml b/fern/apis/prod/openapi/openapi.yaml
@@ -2087,7 +2087,7 @@ components:
         description:
           title: Description
           type: string
-          default: An helpful agent.
+          default: A helpful agent.
         avatar:
           title: Avatar
           type: string
@@ -2239,7 +2239,7 @@ components:
         description:
           title: Description
           type: string
-          default: An helpful tool.
+          default: a helpful tool.
         type:
           title: Type
           type: string

diff --git a/fern/docs.yml b/fern/docs.yml
@@ -44,8 +44,12 @@ navigation:
             path: ./mdx/sdk/workflows.mdx
           - page: Configuring Vector Database
             path: ./mdx/sdk/configure_vector_database.mdx
-          - page: SuperAgent Markup Language
-            path: ./mdx/sdk/saml.mdx
+      - section: SAML
+        contents:
+          - page: Intro
+            path: ./mdx/saml/intro.mdx
+          - page: Structured Outputs
+            path: ./mdx/saml/structured_outputs.mdx
       - section: Installation 
         contents:
           - page: Docker Compose

diff --git a/fern/mdx/apps/image-gen.mdx b/fern/mdx/apps/image-gen.mdx
@@ -65,7 +65,7 @@ With our client setup we can now go ahead and create a `Replicate` tool and corr
 <CodeBlocks>
     <CodeBlock title="Python"> 
         ```python
-        promt = """You are an helpful AI Assistant that can use Playground AI API to generate images.
+        promt = """You are a helpful AI Assistant that can use Playground AI API to generate images.
 
         Follow these steps:
 
@@ -100,7 +100,7 @@ With our client setup we can now go ahead and create a `Replicate` tool and corr
     <CodeBlock title="Javascript"> 
         ```javascript
         prompt = `
-        You are an helpful AI Assistant that can use Playground AI API to generate images.
+        You are a helpful AI Assistant that can use Playground AI API to generate images.
 
         Follow these steps:
 
@@ -185,7 +185,7 @@ Below you will find the full code for this assistant. You can optionaly use the
             base_url="https://api.beta.superagent.sh" # or your local environment
         )
 
-        promt = """You are an helpful AI Assistant that can use Playground AI API to generate images.
+        promt = """You are a helpful AI Assistant that can use Playground AI API to generate images.
 
         Follow these steps:
 
@@ -236,7 +236,7 @@ Below you will find the full code for this assistant. You can optionaly use the
         })
 
         prompt = `
-        You are an helpful AI Assistant that can use Playground AI API to generate images.
+        You are a helpful AI Assistant that can use Playground AI API to generate images.
 
         Follow these steps:
 

diff --git a/fern/mdx/sdk/saml.mdx → fern/mdx/saml/intro.mdx b/fern/mdx/sdk/saml.mdx → fern/mdx/saml/intro.mdx
diff --git a/fern/mdx/saml/structured_outputs.mdx b/fern/mdx/saml/structured_outputs.mdx
@@ -0,0 +1,66 @@
+Besides SDK we support setting output schemas as well. In this page we will show you how to set output schema for forcing LLM to generate output in a specific format.
+Output schema is a JSON object that describes the structure of the output. This can be anything from a simple JSON object to a complex nested structure.
+
+
+Note: We will use `Scraper` tool in our examples. You can get the API Key from [ScrapingBee's website](https://www.scrapingbee.com/)
+
+### Natural Language Schema
+This is a simple and intuitive way to define the output schema. It can be anything.
+
+```yaml
+workflows:
+  - anthropic:
+      llm: claude-3-haiku-20240307
+      name: Structured Assistant
+      tools:
+        - scraper:
+            name: browser
+            use_for: searching the internet
+            metadata:
+              apiKey:  <YOUR_SCRAPINGBEE_API_KEY>
+      prompt: You're a helpful assistant.
+      output_schema: |-
+        [
+          {
+           product_url: Product URL that links to the Amazon page
+           product_name: Name of the product
+           product_price: Price of the product in U.S dollars
+          }
+        ]
+```
+
+
+### JSON Schema
+JSON schema is more powerful than natural language schema. It allows you to define the output schema in a more structured way.
+
+```yaml
+workflows:
+  - anthropic:
+      llm: claude-3-haiku-20240307
+      name: Structured Assistant
+      tools:
+        - scraper:
+            name: browser
+            use_for: searching the internet
+            metadata:
+              apiKey: <YOUR_SCRAPINGBEE_API_KEY>
+      prompt: You're a helpful assistant.
+      output_schema:
+        type: array
+        items:
+          type: object
+          required:
+            - product_url
+            - product_name
+            - product_price
+          properties:
+            product_url:
+              type: string
+            product_name:
+              type: string
+            product_price:
+              type: number
+              description: the price of the product in U.S dollars
+```
+
+For example, you can ask for `I need a list of books about programming from Amazon.`. The scraper tool will scrape the data from Amazon's website and return the data in the format you have defined in the output schema.
diff --git a/fern/mdx/sdk/basic_example.mdx b/fern/mdx/sdk/basic_example.mdx
@@ -145,7 +145,7 @@ By seperating the creation of each object you can reuse, LLMs, Agents or any oth
             "isActive": True,
             "initialMessage": "Hi there! How can I help you?",
             "llmModel": "GPT_3_5_TURBO_16K_0613",
-            "prompt": "You are an helpful AI Assistant", 
+            "prompt": "You are a helpful AI Assistant", 
         })
 
         client.agent.add_llm(agent_id=agent.data.id, llm_id=llm.data.id)
@@ -184,7 +184,7 @@ By seperating the creation of each object you can reuse, LLMs, Agents or any oth
             isActive: true,
             initialMessage: "Hi there! How can I help you?",
             llmModel: "GPT_3_5_TURBO_16K_0613",
-            prompt: "You are an helpful AI Assistant", 
+            prompt: "You are a helpful AI Assistant", 
         })
 
 

diff --git a/fern/mdx/sdk/structured_outputs.mdx b/fern/mdx/sdk/structured_outputs.mdx
@@ -83,7 +83,7 @@ You can force your Assistant to reply using structured outputs. This can be bene
         ```python 
         prediction = client.agent.invoke(
             agent_id=agent.data.id,
-            input="List the top 5 articles on https://news.ycombinator.com."
+            input="List the top 5 articles on https://news.ycombinator.com.",
             enable_streaming=False,
             session_id="my_session_id",
             output_schema="[{title: string, points: number, url: string}]" # Your desired output schema
@@ -103,7 +103,7 @@ You can force your Assistant to reply using structured outputs. This can be bene
     <CodeBlock title="Javascript">
         ```javascript 
         const {data: prediction} = await client.agent.invoke(agent.id, {
-            input: "List the top 5 articles on https://news.ycombinator.com."
+            input: "List the top 5 articles on https://news.ycombinator.com.",
             enableStreaming: false,
             sessionId: "my_session_id",
             outputSchema: "[{title: string, points: number, url: string}]" // Your desired output schema

diff --git a/libs/superagent/app/agents/base.py b/libs/superagent/app/agents/base.py
@@ -69,6 +69,7 @@ async def get_agent(self):
                 agent_id=self.agent_id,
                 session_id=self.session_id,
                 enable_streaming=self.enable_streaming,
+                output_schema=self.output_schema,
                 callbacks=self.callbacks,
                 llm_params=self.llm_params,
                 agent_config=self.agent_config,

diff --git a/libs/superagent/app/agents/langchain.py b/libs/superagent/app/agents/langchain.py
@@ -23,7 +23,8 @@
 from app.tools import TOOL_TYPE_MAPPING, create_pydantic_model_from_object, create_tool
 from app.tools.datasource import DatasourceTool, StructuredDatasourceTool
 from app.utils.llm import LLM_MAPPING
-from prisma.models import LLM, Agent, AgentDatasource, AgentTool
+from lib.prompts import JSON_FORMAT_INSTRUCTIONS
+from prisma.models import LLM, AgentDatasource, AgentTool
 
 DEFAULT_PROMPT = (
     "You are a helpful AI Assistant, answer the users questions to "
@@ -159,10 +160,10 @@ def get_llm_params(self):
             "max_tokens": options.get("max_tokens"),
         }
 
-    async def _get_llm(self, llm: LLM, agent: Agent):
+    async def _get_llm(self, llm: LLM):
         if llm.provider == "OPENAI":
             return ChatOpenAI(
-                model=LLM_MAPPING[agent.llmModel],
+                model=LLM_MAPPING[self.agent_config.llmModel],
                 openai_api_key=llm.apiKey,
                 streaming=self.enable_streaming,
                 callbacks=self.callbacks,
@@ -176,26 +177,16 @@ async def _get_llm(self, llm: LLM, agent: Agent):
                 **self.get_llm_params(),
             )
 
-    async def _get_prompt(self, agent: Agent):
-        base_prompt = agent.prompt or DEFAULT_PROMPT
-        if self.output_schema:
-            content = f"""
-                {base_prompt}\n\n"
-                Always answer using the below output schema. 
-                The output should be formatted as a JSON instance that conforms to the JSON schema below.
-
-                As an example, for the schema {{"properties": {{"foo": {{"title": "Foo", "description": "a list of strings", "type": "array", "items": {{"type": "string"}}}}}}, "required": ["foo"]}}
-                the object {{"foo": ["bar", "baz"]}} is a well-formatted instance of the schema. The object {{"properties": {{"foo": ["bar", "baz"]}}}} is not well-formatted.
+    async def _get_prompt(self):
+        base_prompt = self.agent_config.prompt or DEFAULT_PROMPT
+        content = f"{datetime.datetime.now().strftime('%Y-%m-%d')}"
 
-                Here is the output schema:
-                ```
-                {self.output_schema}
-                ```
-                """
+        if self.output_schema:
+            content += JSON_FORMAT_INSTRUCTIONS.format(
+                base_prompt=base_prompt, output_schema=self.output_schema
+            )
         else:
-            content = f"{base_prompt}"
-
-        content += f"\n\nCurrent date: {datetime.datetime.now().strftime('%Y-%m-%d')}"
+            content += f"{base_prompt}"
 
         return SystemMessage(content=content)
 
@@ -231,14 +222,12 @@ async def _get_memory(
         return memory
 
     async def get_agent(self):
-        llm = await self._get_llm(
-            llm=self.agent_config.llms[0].llm, agent=self.agent_config
-        )
+        llm = await self._get_llm(llm=self.agent_config.llms[0].llm)
         tools = await self._get_tools(
             agent_datasources=self.agent_config.datasources,
             agent_tools=self.agent_config.tools,
         )
-        prompt = await self._get_prompt(agent=self.agent_config)
+        prompt = await self._get_prompt()
         memory = await self._get_memory()
 
         if len(tools) > 0:

diff --git a/libs/superagent/app/agents/llm.py b/libs/superagent/app/agents/llm.py
@@ -1,3 +1,4 @@
+import datetime
 import logging
 from typing import Optional
 
@@ -7,6 +8,7 @@
 from app.utils.callbacks import CustomAsyncIteratorCallbackHandler
 from app.utils.llm import LLM_REVERSE_MAPPING
 from app.utils.prisma import prisma
+from lib.prompts import JSON_FORMAT_INSTRUCTIONS
 from prisma.enums import AgentType, LLMProvider
 from prisma.models import Agent
 
@@ -59,6 +61,12 @@ async def init(self):
         return self
 
 
+DEFAULT_PROMPT = (
+    "You are a helpful AI Assistant, answer the users questions to "
+    "the best of your ability."
+)
+
+
 class LLMAgent(AgentBase):
     def get_llm_params(self):
         llm = self.agent_config.llms[0].llm
@@ -74,10 +82,31 @@ def get_llm_params(self):
             "max_tokens": options.get("max_tokens"),
         }
 
+    async def _get_prompt(self):
+        base_prompt = self.agent_config.prompt or DEFAULT_PROMPT
+        print("OUTPUT SCHEMA", self.output_schema)
+
+        prompt = f"Current date: {datetime.datetime.now().strftime('%Y-%m-%d')}\n"
+
+        if self.output_schema:
+            prompt += f"""
+                {JSON_FORMAT_INSTRUCTIONS.format(
+                    base_prompt=base_prompt, output_schema=self.output_schema
+                )}
+                Always surround the output with "```json```" to ensure proper formatting.
+                """
+        else:
+            prompt = base_prompt
+
+        return prompt
+
     async def get_agent(self):
         enable_streaming = self.enable_streaming
         agent_config = self.agent_config
         session_id = self.session_id
+        model = agent_config.metadata.get("model", "gpt-3.5-turbo-0125")
+        api_key = agent_config.llms[0].llm.apiKey
+        prompt = await self._get_prompt()
 
         class CustomAgentExecutor:
             def __init__(self, llm_agent_instance: LLMAgent, *args, **kwargs):
@@ -102,10 +131,6 @@ async def ainvoke(self, input, *_, **kwargs):
                         input=input
                     )
 
-                model = agent_config.metadata.get("model", "gpt-3.5-turbo-0125")
-                prompt = agent_config.prompt
-                api_key = agent_config.llms[0].llm.apiKey
-
                 if function_calling_res.get("output"):
                     INPUT_TEMPLATE = "{input}\n Context: {context}\n"
                     input = INPUT_TEMPLATE.format(

diff --git a/libs/superagent/app/agents/test.json b/libs/superagent/app/agents/test.json
@@ -0,0 +1,11 @@
+[
+  {
+    "product_url": "https://www.amazon.com/gp/product/B0B12BQ44J",
+    "product_name": "HP 14\" Ultral Light Laptop for Students and Business, Intel Quad-Core N4120, 16GB RAM, 192GB Storage(64GB eMMC+128GB Ghost Manta SD), 1 Year Office 365, Webcam, HDMI, WiFi, USB-A&C, Win 11",
+    "product_price": 799.99
+  },
+  {
+    "product_url": "https://www.amazon.com/gp/product/B0B12BQ44J",
+    "product_name": "HP 14\" Ultral Light Laptop for Students and Business, Intel Quad-Core N4120, 8GB RAM, 192GB Storage(64GB eMMC+128GB Micro SD), 1 Year Office 365, Webcam, HDMI, WiFi, USB-A&C, Win S"
+  }
+]
diff --git a/libs/superagent/app/models/request.py b/libs/superagent/app/models/request.py
@@ -32,7 +32,7 @@ class Agent(BaseModel):
     prompt: Optional[str]
     llmModel: Optional[str]
     llmProvider: Optional[LLMProvider]
-    description: Optional[str] = "An helpful agent."
+    description: Optional[str] = "a helpful agent."
     avatar: Optional[str]
     type: Optional[AgentType] = AgentType.SUPERAGENT
     parameters: Optional[OpenAiAssistantParameters]
@@ -116,7 +116,7 @@ class DatasourceUpdate(BaseModel):
 
 class Tool(BaseModel):
     name: str
-    description: Optional[str] = "An helpful tool."
+    description: Optional[str] = "a helpful tool."
     type: str
     metadata: Optional[Dict[Any, Any]]
     returnDirect: Optional[bool] = False

diff --git a/libs/superagent/app/tools/flow.py b/libs/superagent/app/tools/flow.py
@@ -84,7 +84,7 @@ async def generate_route(function_schema: Dict[str, Any]) -> str:
         model="openrouter/mistralai/mixtral-8x7b-instruct",
         api_key=config("OPENROUTER_API_KEY"),
         messages=[
-            {"role": "system", "content": "You are an helpful assistant."},
+            {"role": "system", "content": "You are a helpful assistant."},
             {
                 "role": "user",
                 "content": prompt,

diff --git a/libs/superagent/app/tools/prompts.py b/libs/superagent/app/tools/prompts.py
@@ -42,7 +42,7 @@ def create_function_response_prompt(input: str, context: str) -> str:
     """
 
     prompt = (
-        "You are an helpful AI Assistant, answer the question by "
+        "You are a helpful AI Assistant, answer the question by "
         "providing the most suitable response based on the context provided.\n\n"
         f"Input: {input}\n\n"
         f"Context:\n{context}"

diff --git a/libs/superagent/app/workflows/base.py b/libs/superagent/app/workflows/base.py
@@ -9,18 +9,6 @@
 from app.agents.base import AgentBase
 from app.utils.callbacks import CustomAsyncIteratorCallbackHandler
 
-# Adapted from https://github.com/langchain-ai/langchain/blob/d1a2e194c376f241116bf8e520f1a9bb297cdf3a/libs/core/langchain_core/output_parsers/format_instructions.py
-JSON_FORMAT_INSTRUCTIONS = """The output should be formatted as a JSON instance that conforms to the JSON schema below.
-
-As an example, for the schema {{"properties": {{"foo": {{"title": "Foo", "description": "a list of strings", "type": "array", "items": {{"type": "string"}}}}}}, "required": ["foo"]}}
-the object {{"foo": ["bar", "baz"]}} is a well-formatted instance of the schema. The object {{"properties": {{"foo": ["bar", "baz"]}}}} is not well-formatted.
-
-Here is the output schema:
-```
-{schema}
-```
-"""
-
 
 class WorkflowBase:
     def __init__(
@@ -69,6 +57,7 @@ async def arun(self, input: Any):
                     "callbacks": self.callbacks[stepIndex],
                 },
             )
+            print("agent_response", agent_response)
             if output_schema:
                 # TODO: throw error if output is not valid
                 json_parser = SimpleJsonOutputParser()