The modern corporate world is filled with microservices and APIs, which are developed by various teams at an extremely fast pace. Despite the popularity of tools like FastAPI, which allows generating specifications based on Pydantic, one of the main challenges startups face is the lack of time to write quality documentation. However, if part of the code is already being generated automatically, it might be possible to reduce the amount of manual work. In this article, I want to share my experience, where I decided to conduct an experiment: combining OpenAPI with LLM and observing how it affects the generation and accessibility of documentation. Using Retrieval Augmented Generation (RAG), ChromaDB, and Langchain, I built a small chat prototype that integrates with OpenAPI Specification. This article is a detailed guide for those who want to learn how the experiment went and what steps I took to create a system that would allow for quickly finding answers to questions such as "How to create a user via API?" or "How to get a list of users?" Retrieval Augmented Generation Retrieval Augmented Generation (RAG) is an approach that combines generative models with a search mechanism to retrieve information from external data sources. RAG enhances the quality and accuracy of responses, as the model utilizes not only its trained parameters but also up-to-date data retrieved from information repositories, such as databases or vector stores. The main steps in implementing a RAG system are as follows: Creating a "knowledge base": The information that the system will work with is divided into small chunks. These chunks then go through an encoding process—transforming the text into vector representations, which are stored in a vector database for quick access. In our case, we will manually divide the OpenAPI specification into chunks and store their vector representations in ChromaDB. Retrieving data during a query: When a user sends a query, the system employs a search mechanism to extract the relevant chunks of information from the vector database. These data are then sent to the generative model to formulate a response. To retrieve the data, we will use Langchain. Write the OpenAPI Splitter Langchain already offers numerous splitters. Their purpose is to divide the text in a way that allows semantically similar parts to be found during vector search and passed into the context. In our case, we need to pass the entire context about an HTTP request from the documentation. To achieve this, it's sufficient to simply split our JSON into parts for each path. def get_openapi_spec_paths(specification: dict) -> dict: paths = [] for p in specification["paths"]: for m in specification["paths"][p]: path = specification["paths"][p][m] path["method"] = m path["path"] = p paths.append(path) return paths Embed Paths After splitting our OpenAPI specification into chunks, we need to obtain their vector representation and load them into the ChromaDB vector store. Later, before making a request to the LLM, we will perform a vector search for relevant information and add the context. import json from langchain.docstore.document import Document from langchain_openai import OpenAIEmbeddings from langchain_chroma import Chroma specification = get_openapi_spec(url) paths = get_openapi_spec_paths(specification) dumped_paths = dump_openapi_spec_to_chroma_docs(paths) for p in paths: dumped_paths.append( Document( page_content=json.dumps(p), metadata={"source": "local"}) ) embeddings = OpenAIEmbeddings( model="text-embedding-ada-002" ) Chroma.from_documents( documents=dumped_paths, embedding=embeddings, persist_directory="data", collection_name="spec" ) In this example, we store the data locally in the 'data' directory, but Chroma can operate both in-memory and as a standalone instance. This can be found in the documentation. Retrieving and Chaining Spec We are ready to make the first request. I have uploaded an example of a simple API about dogs, which I wrote using FastAPI, and included the link in the script. For the model, we will use ChatGPT-4-turbo. embeddings = OpenAIEmbeddings(model=settings.OPENAI_API_EMBEDDINGS_MODEL) llm = ChatOpenAI(api_key=settings.OPENAI_API_KEY, model=settings.OPEN_API_MODEL) chroma_db = Chroma( persist_directory="data", embedding_function=embeddings, collection_name="spec", ) retriever = chroma_db.as_retriever() prompt = PromptTemplate.from_template( """ System: You are an assistant that converts OpenAPI JSON specs into neatly structured, human-readable text with summary, description, tags, produces, responses, parameters, method, and curl examples. EXAMPLE ANSWER: ### GET /dogs **Summary**: Retrieve a list of dogs **Description**: Get a filtered list of dogs based on query parameters such as breed, age, or size. **Tags**: Dogs **Produces**: application/json **Parameters**: - **Query**: - `breed` (string, optional): Filter by dog breed. - `age` (integer, optional): Filter by dog age. - `size` (string, optional): Filter by dog size (small, medium, large). **Method**: GET **Curl Example**: curl -X GET "https://api.example.com/dogs?breed=labrador&size=medium" -H "Authorization: Bearer <your_token>" Now, format the following OpenAPI spec: {context} {question} """ ) def format_docs(docs): return \n \\n".join(doc.page_content for doc in docs) chain = ( {"context": retriever | format_docs, "question": RunnablePassthrough()} | prompt | llm | StrOutputParser() ) chain.invoke(("How to get list of dogs?") Llama Example If you're not ready to use OpenAPI, for example, due to security concerns or if you don't have an API key, there is always the option to run a local instance on Llama. Download ollama and pull the model. ollama pull llama3.1 Code example. from langchain_ollama import OllamaLLM llm = OllamaLLM(model="llama3.1:8b") Let’s Streamlit Let’s add Streamlit to quickly create a web interface for our chat. Streamlit is an open-source Python framework for data scientists and AI/ML engineers to deliver dynamic data apps - in only a few lines of code. Make a file chat.py. import streamlit as st if "messages" not in st.session_state: st.session_state.messages = [] for message in st.session_state.messages: with st.chat_message(message["role"]): st.markdown(message["content"]) if prompt := st.chat_input("Enter your message."): st.session_state.messages.append({"role": "user", "content": prompt}) with st.chat_message("user"): st.markdown(prompt) with st.chat_message("assistant"): if (len(st.session_state.messages)): r = chat_with_model(st.session_state.messages[-1]["content"]) response = st.write(r) st.session_state.messages.append({"role": "assistant", "content": response}) Run it streamline run chat.py Run on Fake Spec Before running, I generated a simple API using ChatGPT. It was easy. from fastapi import FastAPI, HTTPException from pydantic import BaseModel from typing import List, Optional app = FastAPI() # In-memory storage for dogs dogs_db = {} # Pydantic model for a dog class Dog(BaseModel): name: str breed: str age: int class DogUpdate(BaseModel): name: Optional[str] = None breed: Optional[str] = None age: Optional[int] = None @app.get("/dogs", response_model=List[Dog]) def get_dogs(breed: Optional[str] = None, min_age: Optional[int] = None, max_age: Optional[int] = None): """ Get a list of all dogs. Optionally filter by breed, minimum age, and maximum age. - **breed**: Filter by dog breed. - **min_age**: Minimum age of the dogs to retrieve. - **max_age**: Maximum age of the dogs to retrieve. """ filtered_dogs = [dog for dog in dogs_db.values()] if breed: filtered_dogs = [dog for dog in filtered_dogs if dog.breed == breed] if min_age is not None: filtered_dogs = [dog for dog in filtered_dogs if dog.age >= min_age] if max_age is not None: filtered_dogs = [dog for dog in filtered_dogs if dog.age <= max_age] return filtered_dogs @app.post("/dogs", response_model=Dog) def add_dog(dog: Dog): """ Add a new dog to the database. - **name**: The name of the dog. - **breed**: The breed of the dog. - **age**: The age of the dog. """ dog_id = len(dogs_db) + 1 dogs_db[dog_id] = dog return dog @app.get("/dogs/{dog_id}", response_model=Dog) def get_dog(dog_id: int): """ Get a specific dog by its ID. - **dog_id**: The ID of the dog to retrieve. """ dog = dogs_db.get(dog_id) if not dog: raise HTTPException(status_code=404, detail="Dog not found") return dog @app.put("/dogs/{dog_id}", response_model=Dog) def update_dog(dog_id: int, dog_update: DogUpdate): """ Update a dog's information by its ID. Only the fields provided in the request body will be updated. - **dog_id**: The ID of the dog to update. - **name**: The new name of the dog (optional). - **breed**: The new breed of the dog (optional). - **age**: The new age of the dog (optional). """ dog = dogs_db.get(dog_id) if not dog: raise HTTPException(status_code=404, detail="Dog not found") if dog_update.name is not None: dog.name = dog_update.name if dog_update.breed is not None: dog.breed = dog_update.breed if dog_update.age is not None: dog.age = dog_update.age dogs_db[dog_id] = dog return dog Let's see what we got in the chat? Wow, it works. As a result, I spent 2 hours playing around with working specifications and was amazed by the generation. Conclusion In this experiment, I demonstrated how the use of Retrieval Augmented Generation (RAG) can significantly simplify working with API documentation and specifications. With RAG and tools like Langchain and ChromaDB, we created a system that effectively responds to queries about OpenAPI, reducing the time spent searching for necessary information. The integration of LLM with vector search opens new possibilities for automating documentation, which is especially beneficial for startups and teams with limited resources. In the following articles, there will be even more experiments. If you found this experiment helpful, please support the project with a star on GitHub! RAG can become an important tool in automating documentation work. Here is the link to the repository for local project deployment. The modern corporate world is filled with microservices and APIs, which are developed by various teams at an extremely fast pace. Despite the popularity of tools like FastAPI, which allows generating specifications based on Pydantic, one of the main challenges startups face is the lack of time to write quality documentation. However, if part of the code is already being generated automatically, it might be possible to reduce the amount of manual work. In this article, I want to share my experience, where I decided to conduct an experiment: combining OpenAPI with LLM and observing how it affects the generation and accessibility of documentation. Using Retrieval Augmented Generation (RAG), ChromaDB, and Langchain, I built a small chat prototype that integrates with OpenAPI Specification. This article is a detailed guide for those who want to learn how the experiment went and what steps I took to create a system that would allow for quickly finding answers to questions such as "How to create a user via API?" or "How to get a list of users?" Retrieval Augmented Generation Retrieval Augmented Generation (RAG) is an approach that combines generative models with a search mechanism to retrieve information from external data sources. RAG enhances the quality and accuracy of responses, as the model utilizes not only its trained parameters but also up-to-date data retrieved from information repositories, such as databases or vector stores. The main steps in implementing a RAG system are as follows: Creating a "knowledge base": The information that the system will work with is divided into small chunks. These chunks then go through an encoding process—transforming the text into vector representations, which are stored in a vector database for quick access. In our case, we will manually divide the OpenAPI specification into chunks and store their vector representations in ChromaDB. Retrieving data during a query: When a user sends a query, the system employs a search mechanism to extract the relevant chunks of information from the vector database. These data are then sent to the generative model to formulate a response. To retrieve the data, we will use Langchain. Creating a "knowledge base": The information that the system will work with is divided into small chunks. These chunks then go through an encoding process—transforming the text into vector representations, which are stored in a vector database for quick access. In our case, we will manually divide the OpenAPI specification into chunks and store their vector representations in ChromaDB. Creating a "knowledge base" : The information that the system will work with is divided into small chunks. These chunks then go through an encoding process—transforming the text into vector representations, which are stored in a vector database for quick access. In our case, we will manually divide the OpenAPI specification into chunks and store their vector representations in ChromaDB. Creating a "knowledge base" Retrieving data during a query: When a user sends a query, the system employs a search mechanism to extract the relevant chunks of information from the vector database. These data are then sent to the generative model to formulate a response. To retrieve the data, we will use Langchain. Retrieving data during a query : When a user sends a query, the system employs a search mechanism to extract the relevant chunks of information from the vector database. These data are then sent to the generative model to formulate a response. To retrieve the data, we will use Langchain. Retrieving data during a query Write the OpenAPI Splitter Langchain already offers numerous splitters. Their purpose is to divide the text in a way that allows semantically similar parts to be found during vector search and passed into the context. In our case, we need to pass the entire context about an HTTP request from the documentation. To achieve this, it's sufficient to simply split our JSON into parts for each path. def get_openapi_spec_paths(specification: dict) -> dict: paths = [] for p in specification["paths"]: for m in specification["paths"][p]: path = specification["paths"][p][m] path["method"] = m path["path"] = p paths.append(path) return paths def get_openapi_spec_paths(specification: dict) -> dict: paths = [] for p in specification["paths"]: for m in specification["paths"][p]: path = specification["paths"][p][m] path["method"] = m path["path"] = p paths.append(path) return paths Embed Paths After splitting our OpenAPI specification into chunks, we need to obtain their vector representation and load them into the ChromaDB vector store. Later, before making a request to the LLM, we will perform a vector search for relevant information and add the context. import json from langchain.docstore.document import Document from langchain_openai import OpenAIEmbeddings from langchain_chroma import Chroma specification = get_openapi_spec(url) paths = get_openapi_spec_paths(specification) dumped_paths = dump_openapi_spec_to_chroma_docs(paths) for p in paths: dumped_paths.append( Document( page_content=json.dumps(p), metadata={"source": "local"}) ) embeddings = OpenAIEmbeddings( model="text-embedding-ada-002" ) Chroma.from_documents( documents=dumped_paths, embedding=embeddings, persist_directory="data", collection_name="spec" ) import json from langchain.docstore.document import Document from langchain_openai import OpenAIEmbeddings from langchain_chroma import Chroma specification = get_openapi_spec(url) paths = get_openapi_spec_paths(specification) dumped_paths = dump_openapi_spec_to_chroma_docs(paths) for p in paths: dumped_paths.append( Document( page_content=json.dumps(p), metadata={"source": "local"}) ) embeddings = OpenAIEmbeddings( model="text-embedding-ada-002" ) Chroma.from_documents( documents=dumped_paths, embedding=embeddings, persist_directory="data", collection_name="spec" ) In this example, we store the data locally in the 'data' directory, but Chroma can operate both in-memory and as a standalone instance. This can be found in the documentation. Retrieving and Chaining Spec We are ready to make the first request. I have uploaded an example of a simple API about dogs, which I wrote using FastAPI, and included the link in the script. For the model, we will use ChatGPT-4-turbo. embeddings = OpenAIEmbeddings(model=settings.OPENAI_API_EMBEDDINGS_MODEL) llm = ChatOpenAI(api_key=settings.OPENAI_API_KEY, model=settings.OPEN_API_MODEL) chroma_db = Chroma( persist_directory="data", embedding_function=embeddings, collection_name="spec", ) retriever = chroma_db.as_retriever() prompt = PromptTemplate.from_template( """ System: You are an assistant that converts OpenAPI JSON specs into neatly structured, human-readable text with summary, description, tags, produces, responses, parameters, method, and curl examples. EXAMPLE ANSWER: ### GET /dogs **Summary**: Retrieve a list of dogs **Description**: Get a filtered list of dogs based on query parameters such as breed, age, or size. **Tags**: Dogs **Produces**: application/json **Parameters**: - **Query**: - `breed` (string, optional): Filter by dog breed. - `age` (integer, optional): Filter by dog age. - `size` (string, optional): Filter by dog size (small, medium, large). **Method**: GET **Curl Example**: curl -X GET "https://api.example.com/dogs?breed=labrador&size=medium" -H "Authorization: Bearer <your_token>" Now, format the following OpenAPI spec: {context} {question} """ ) def format_docs(docs): return \n \\n".join(doc.page_content for doc in docs) chain = ( {"context": retriever | format_docs, "question": RunnablePassthrough()} | prompt | llm | StrOutputParser() ) chain.invoke(("How to get list of dogs?") embeddings = OpenAIEmbeddings(model=settings.OPENAI_API_EMBEDDINGS_MODEL) llm = ChatOpenAI(api_key=settings.OPENAI_API_KEY, model=settings.OPEN_API_MODEL) chroma_db = Chroma( persist_directory="data", embedding_function=embeddings, collection_name="spec", ) retriever = chroma_db.as_retriever() prompt = PromptTemplate.from_template( """ System: You are an assistant that converts OpenAPI JSON specs into neatly structured, human-readable text with summary, description, tags, produces, responses, parameters, method, and curl examples. EXAMPLE ANSWER: ### GET /dogs **Summary**: Retrieve a list of dogs **Description**: Get a filtered list of dogs based on query parameters such as breed, age, or size. **Tags**: Dogs **Produces**: application/json **Parameters**: - **Query**: - `breed` (string, optional): Filter by dog breed. - `age` (integer, optional): Filter by dog age. - `size` (string, optional): Filter by dog size (small, medium, large). **Method**: GET **Curl Example**: curl -X GET "https://api.example.com/dogs?breed=labrador&size=medium" -H "Authorization: Bearer <your_token>" Now, format the following OpenAPI spec: {context} {question} """ ) def format_docs(docs): return \n \\n".join(doc.page_content for doc in docs) chain = ( {"context": retriever | format_docs, "question": RunnablePassthrough()} | prompt | llm | StrOutputParser() ) chain.invoke(("How to get list of dogs?") Llama Example If you're not ready to use OpenAPI, for example, due to security concerns or if you don't have an API key, there is always the option to run a local instance on Llama. Download ollama and pull the model. ollama pull llama3.1 ollama pull llama3.1 Code example. from langchain_ollama import OllamaLLM llm = OllamaLLM(model="llama3.1:8b") from langchain_ollama import OllamaLLM llm = OllamaLLM(model="llama3.1:8b") Let’s Streamlit Let’s add Streamlit to quickly create a web interface for our chat. Streamlit Streamlit is an open-source Python framework for data scientists and AI/ML engineers to deliver dynamic data apps - in only a few lines of code. an open-source Python framework for data scientists and AI/ML engineers to deliver dynamic data apps Make a file chat.py . chat.py import streamlit as st if "messages" not in st.session_state: st.session_state.messages = [] for message in st.session_state.messages: with st.chat_message(message["role"]): st.markdown(message["content"]) if prompt := st.chat_input("Enter your message."): st.session_state.messages.append({"role": "user", "content": prompt}) with st.chat_message("user"): st.markdown(prompt) with st.chat_message("assistant"): if (len(st.session_state.messages)): r = chat_with_model(st.session_state.messages[-1]["content"]) response = st.write(r) st.session_state.messages.append({"role": "assistant", "content": response}) import streamlit as st if "messages" not in st.session_state: st.session_state.messages = [] for message in st.session_state.messages: with st.chat_message(message["role"]): st.markdown(message["content"]) if prompt := st.chat_input("Enter your message."): st.session_state.messages.append({"role": "user", "content": prompt}) with st.chat_message("user"): st.markdown(prompt) with st.chat_message("assistant"): if (len(st.session_state.messages)): r = chat_with_model(st.session_state.messages[-1]["content"]) response = st.write(r) st.session_state.messages.append({"role": "assistant", "content": response}) Run it streamline run chat.py streamline run chat.py Run on Fake Spec Before running, I generated a simple API using ChatGPT. It was easy. from fastapi import FastAPI, HTTPException from pydantic import BaseModel from typing import List, Optional app = FastAPI() # In-memory storage for dogs dogs_db = {} # Pydantic model for a dog class Dog(BaseModel): name: str breed: str age: int class DogUpdate(BaseModel): name: Optional[str] = None breed: Optional[str] = None age: Optional[int] = None @app.get("/dogs", response_model=List[Dog]) def get_dogs(breed: Optional[str] = None, min_age: Optional[int] = None, max_age: Optional[int] = None): """ Get a list of all dogs. Optionally filter by breed, minimum age, and maximum age. - **breed**: Filter by dog breed. - **min_age**: Minimum age of the dogs to retrieve. - **max_age**: Maximum age of the dogs to retrieve. """ filtered_dogs = [dog for dog in dogs_db.values()] if breed: filtered_dogs = [dog for dog in filtered_dogs if dog.breed == breed] if min_age is not None: filtered_dogs = [dog for dog in filtered_dogs if dog.age >= min_age] if max_age is not None: filtered_dogs = [dog for dog in filtered_dogs if dog.age <= max_age] return filtered_dogs @app.post("/dogs", response_model=Dog) def add_dog(dog: Dog): """ Add a new dog to the database. - **name**: The name of the dog. - **breed**: The breed of the dog. - **age**: The age of the dog. """ dog_id = len(dogs_db) + 1 dogs_db[dog_id] = dog return dog @app.get("/dogs/{dog_id}", response_model=Dog) def get_dog(dog_id: int): """ Get a specific dog by its ID. - **dog_id**: The ID of the dog to retrieve. """ dog = dogs_db.get(dog_id) if not dog: raise HTTPException(status_code=404, detail="Dog not found") return dog @app.put("/dogs/{dog_id}", response_model=Dog) def update_dog(dog_id: int, dog_update: DogUpdate): """ Update a dog's information by its ID. Only the fields provided in the request body will be updated. - **dog_id**: The ID of the dog to update. - **name**: The new name of the dog (optional). - **breed**: The new breed of the dog (optional). - **age**: The new age of the dog (optional). """ dog = dogs_db.get(dog_id) if not dog: raise HTTPException(status_code=404, detail="Dog not found") if dog_update.name is not None: dog.name = dog_update.name if dog_update.breed is not None: dog.breed = dog_update.breed if dog_update.age is not None: dog.age = dog_update.age dogs_db[dog_id] = dog return dog from fastapi import FastAPI, HTTPException from pydantic import BaseModel from typing import List, Optional app = FastAPI() # In-memory storage for dogs dogs_db = {} # Pydantic model for a dog class Dog(BaseModel): name: str breed: str age: int class DogUpdate(BaseModel): name: Optional[str] = None breed: Optional[str] = None age: Optional[int] = None @app.get("/dogs", response_model=List[Dog]) def get_dogs(breed: Optional[str] = None, min_age: Optional[int] = None, max_age: Optional[int] = None): """ Get a list of all dogs. Optionally filter by breed, minimum age, and maximum age. - **breed**: Filter by dog breed. - **min_age**: Minimum age of the dogs to retrieve. - **max_age**: Maximum age of the dogs to retrieve. """ filtered_dogs = [dog for dog in dogs_db.values()] if breed: filtered_dogs = [dog for dog in filtered_dogs if dog.breed == breed] if min_age is not None: filtered_dogs = [dog for dog in filtered_dogs if dog.age >= min_age] if max_age is not None: filtered_dogs = [dog for dog in filtered_dogs if dog.age <= max_age] return filtered_dogs @app.post("/dogs", response_model=Dog) def add_dog(dog: Dog): """ Add a new dog to the database. - **name**: The name of the dog. - **breed**: The breed of the dog. - **age**: The age of the dog. """ dog_id = len(dogs_db) + 1 dogs_db[dog_id] = dog return dog @app.get("/dogs/{dog_id}", response_model=Dog) def get_dog(dog_id: int): """ Get a specific dog by its ID. - **dog_id**: The ID of the dog to retrieve. """ dog = dogs_db.get(dog_id) if not dog: raise HTTPException(status_code=404, detail="Dog not found") return dog @app.put("/dogs/{dog_id}", response_model=Dog) def update_dog(dog_id: int, dog_update: DogUpdate): """ Update a dog's information by its ID. Only the fields provided in the request body will be updated. - **dog_id**: The ID of the dog to update. - **name**: The new name of the dog (optional). - **breed**: The new breed of the dog (optional). - **age**: The new age of the dog (optional). """ dog = dogs_db.get(dog_id) if not dog: raise HTTPException(status_code=404, detail="Dog not found") if dog_update.name is not None: dog.name = dog_update.name if dog_update.breed is not None: dog.breed = dog_update.breed if dog_update.age is not None: dog.age = dog_update.age dogs_db[dog_id] = dog return dog Let's see what we got in the chat? Wow, it works. As a result, I spent 2 hours playing around with working specifications and was amazed by the generation. Conclusion In this experiment, I demonstrated how the use of Retrieval Augmented Generation (RAG) can significantly simplify working with API documentation and specifications. With RAG and tools like Langchain and ChromaDB, we created a system that effectively responds to queries about OpenAPI, reducing the time spent searching for necessary information. The integration of LLM with vector search opens new possibilities for automating documentation, which is especially beneficial for startups and teams with limited resources. In the following articles, there will be even more experiments. If you found this experiment helpful, please support the project with a star on GitHub! RAG can become an important tool in automating documentation work. Here is the link to the repository for local project deployment. Here is the link to the repository for local project deployment.
