Ask Your Instance: Use Chat GPT To Understand Code on Your ServiceNow Instance
In this tutorial I will show you how to use Chat GPT on a collection of script files from your ServiceNow instance. This allows you to chat with the code that is running on your instance.
You can follow along with the video, or with the blog post below.
Prequisites (for following along)
- An environment to run an iPython notebook.
- A ServiceNow instance to experiment with.
- An OpenAI API (opens in a new tab) key.
- An ActiveLoop (opens in a new tab) token.
Step 0: Setup
Getting your developer instance password
If you want to try this out on your own personal development instance, you may want to get your instance' username and password handy.
To get your developer instance password go to developer.servicenow.com and click on the menu on the top right corner. Then select Manage Instance Password to see your developer instance password.
Setting up an iPython Notebook
Most of the LLM tooling is python-based and one convenient way of working with Python is through an iPython notebook. Google offers a hosted service to run iPython Notebook called Google Colabs. I prefer develop locally using the iPython functionality within VSCode using the Jupyter (opens in a new tab), Python (opens in a new tab) and Pylint (opens in a new tab) extensions.
(If you use Google Colabs you don't need to worry about virtual environments, but you do need to preface all your shell commands with a !
to run them in the shell.)
If you go with a local iPython Notebook, read on.
In a new folder type in the following to create a new virtual environment:
python3 -m venv .venv
Follow that by activating the virtual environment:
source .venv/bin/activate
On windows this would be:
.venv\Scripts\activate
Then select the newly created venv from the interpreter dropdown in VSCode if its not selected already.
To install all the needed libraries we need to start off by running the following (inside our virtual environment):
pip install -qU --upgrade langchain 'deeplake[enterprise]' openai tiktoken pysnc
Step 1: Getting your ServiceNow instance code
Now let's get the code from our ServiceNow instance. For this we'll use the amazing PySNC (opens in a new tab) library. This library allows you to easily interact with your ServiceNow instance from Python.
import os
import pysnc
# Create the directory if it does not exist
os.makedirs('data/sys_script_include', exist_ok=True)
client = pysnc.ServiceNowClient('dev168935', ('admin', getpass.getpass("Dev Instance Password:")))
gr = client.GlideRecord('sys_script_include')
gr.add_query('sys_package', '16ce0f75e8e1211076da10591ad28708')
gr.query()
for i, r in enumerate(gr):
with open(f'data/sys_script_include/{r.name}.js', 'w') as f:
# If r.script is None, replace it with an empty string
name = r.script.get_value() if r.script is not None else ''
# Write the script to the file
f.write(name)
print(f"Finished writing file for record {i}: {r.name}.js")
Step 2: Creating embeddings
Now let's get ready to connect with OpenAI and ActiveLoop AI:
import os
import getpass
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.vectorstores import DeepLake
os.environ["OPENAI_API_KEY"] = getpass.getpass("OpenAI API Key:")
activeloop_token = getpass.getpass("Activeloop Token:")
os.environ["ACTIVELOOP_TOKEN"] = activeloop_token
For the embeddings let's use the Open AI embeddings:
embeddings = OpenAIEmbeddings(disallowed_special=())
Step 3: Chunking the files
Now let's load and split the files using TextLoader
(opens in a new tab) and its load_and_split method:
import os
from langchain.document_loaders import TextLoader
root_dir = 'data/sys_script_include'
docs = []
# Go through each folder
for dirpath, dirnames, filenames in os.walk(root_dir):
# Go through each file
for file in filenames:
try:
# Load up the file as a doc and split
loader = TextLoader(os.path.join(dirpath, file), encoding='utf-8')
docs.extend(loader.load_and_split())
except Exception as e:
pass
Let's confirm the files were split correctly:
print (f"You have {len(docs)} documents\n")
print ("------ Start Document -----")
print (docs[0].page_content[:300])
This should return something like this:
You have 35 documents
------ Start Document -----
function Schema() {}
Schema.fromTable = function fromTable(table, fields) {
var tableSchema = {};
var tableDescriptor = GlideTableDescriptor.get(table);
var glideTableSchema = tableDescriptor.getSchema();
if (glideTableSchema.isEmpty()) {
NiceError.raise("Unknown table: '" + table + "'");
Step 4: Adding the chunks to the vector store
from langchain.text_splitter import CharacterTextSplitter
text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
texts = text_splitter.split_documents(docs)
Now let's add these chunks to our vector store:
from langchain.vectorstores import DeepLake
username = "jessems" # replace with your username from app.activeloop.ai
db = DeepLake(
dataset_path=f"hub://{username}/glidequery",
embedding_function=embeddings,
)
db.add_documents(texts)
Your Deep Lake dataset has been successfully created!
-Dataset(path='hub://jessems/glidequery', tensors=['embedding', 'id', 'metadata', 'text'])
tensor htype shape dtype compression
------- ------- ------- ------- -------
embedding embedding (138, 1536) float32 None
id text (138, 1) str None
metadata json (138, 1) str None
text text (138, 1) str None
['6599c136-221f-11ee-a991-0242ac1c000c',
'6599c2c6-221f-11ee-a991-0242ac1c000c',
...
Now let's load the data lake:
db = DeepLake(
dataset_path="hub://jessems/glidequery",
read_only=True,
embedding_function=embeddings,
)
Step 5: Creating the retriever
Let's define a retriever:
retriever = db.as_retriever()
retriever.search_kwargs["distance_metric"] = "cos"
retriever.search_kwargs["fetch_k"] = 100
retriever.search_kwargs["maximal_marginal_relevance"] = True
retriever.search_kwargs["k"] = 10
Step 6: Creating the chatbot
I've customized this a bit in relation to the original Langchain tutorial so that Open AI returns a markdown response and includes plenty of code snippets.
from langchain.chat_models import ChatOpenAI
from langchain.chains import ConversationalRetrievalChain
from langchain.prompts.prompt import PromptTemplate
from langchain.prompts.chat import (
ChatPromptTemplate,
SystemMessagePromptTemplate,
HumanMessagePromptTemplate,
)
general_system_template = r"""
Your are a professional ServiceNow dveloper. Give a detailed answers aimed at other programmers. Start your explanations off in simple terms. Respond with markdown. Include code snippets if appropriate. If you don't know the answer, simply say you don't know.
----
{context}
----
"""
general_user_template = "Question:```{question}```"
messages = [
SystemMessagePromptTemplate.from_template(general_system_template),
HumanMessagePromptTemplate.from_template(general_user_template)
]
qa_prompt = ChatPromptTemplate.from_messages( messages )
model = ChatOpenAI(model_name="gpt-4")
qa = ConversationalRetrievalChain.from_llm(model, retriever=retriever, combine_docs_chain_kwargs={"prompt": qa_prompt})
Create a function we can use to ask questions:
chat_history = []
def ask(question, chat_history):
result = qa({"question": question, "chat_history": chat_history})
chat_history.append((question, result["answer"]))
with open('answers.md', 'a') as f:
f.write(f"**Question**: {question} \n\n")
f.write(f"**Answer**: {result['answer']} \n\n")
Step 7: Asking questions
Then ask your questions and watch them appear inside answers.md in markdown.
ask('How does GlideQuery work?', chat_history)
**Question**: How does GlideQuery work?
**Answer**: `GlideQuery` is a class in ServiceNow that provides a more modern and flexible way to perform database operations compared to the traditional `GlideRecord` and `GlideAggregate` APIs. It allows developers to build and execute database queries in a more intuitive and succinct manner, while also providing more advanced and powerful query features.
Here's a simple breakdown of how it works:
1. **Initialization**: You create a new `GlideQuery` object by passing the table name to its constructor:
```javascript
var query = new GlideQuery("sys_user")
```
2. **Building the query**: You can then chain methods onto the `GlideQuery` object to build your query. Each method corresponds to a specific database operation or condition. For example, you can add filter conditions using the `where()` method:
```javascript
query.where("active", true)
```
You can also limit the number of records returned using the `limit()` method:
```javascript
query.limit(10)
```
3. **Executing the query**: You can execute the query and retrieve the results using the `toArray()` or `toGlideRecord()` methods:
```javascript
var results = query.toArray()
```
```javascript
var glideRecord = query.toGlideRecord()
glideRecord.query()
while (glideRecord.next()) {
// Do something with the glideRecord
}
```
Under the hood, the `GlideQuery` class maintains an internal representation of the query, which is updated each time you call a method to modify the query. This internal representation is used to generate the actual SQL query that is executed against the database when you call `toArray()` or `toGlideRecord()`.
In addition to the standard query operations provided by `GlideRecord`, `GlideQuery` also supports more advanced features like subqueries, complex join operations, and aggregation functions.
ask('How is GlideQuery implemented that it allows for you to chain methods on top of each other? ', chat_history)
**Question**: How is GlideQuery implemented that it allows for you to chain methods on top of each other?
**Answer**: GlideQuery is designed to allow method chaining by returning a new GlideQuery object after each method call. This is made possible because each method call adds a new object to the 'plan' array representing the action to be taken, then returns a new GlideQuery with the updated 'plan'.
Let's take a look at the 'withAcls()' method as an example:
```javascript
GlideQuery.prototype.withAcls = function withAcls() {
return new GlideQuery(
this.table,
this.plan.concat({
type: "withAcls",
})
)
}
```
In the code above, a new GlideQuery object is created and returned. The new GlideQuery has the same 'table' as the current GlideQuery, and a 'plan' array that includes all current plan steps plus a new step of the type 'withAcls'. This returned object can then be used for further method chaining.
So if you have a GlideQuery, you can chain methods like this:
```javascript
var users = new GlideQuery("sys_user")
.withAcls()
.limit(20)
.orderByDesc("first_name")
.select("first_name")
.toArray(100)
```
Each method call in the chain adds a new step to the 'plan' and returns a new GlideQuery object, allowing for the next method in the chain to be called.
ask("What about the toArray() method, I don't understand why we need that. What does it do?", chat_history)
**Question**: What about the toArray() method, I don't understand why we need that. What does it do?
**Answer**: The `toArray()` method in GlideQuery is used to fetch the records from the ServiceNow database and returns an array of JavaScript objects representing the records. The objects in the array have properties and values corresponding to the fields and values of the records.
This method is terminal, meaning it executes the GlideQuery, fetches the records, and transforms them into an array of JavaScript objects. After calling `toArray()`, you cannot add more operations to the GlideQuery.
Here's a simple example:
```javascript
var users = new GlideQuery("sys_user")
.select("name", "email")
.limit(10)
.toArray()
users.forEach(function (user) {
gs.info(user.name + ": " + user.email)
})
```
In this example, `users` is an array of objects. Each object represents a user record and has `name` and `email` properties. The `forEach` loop is used to print the name and email of each user.
Behind the scenes, GlideQuery converts each GlideRecord it fetches into a JavaScript object. The returned JavaScript objects are plain data objects, they do not have the methods a GlideRecord has. If you need to call GlideRecord methods on the records, you should use `forEach()`, `map()`, or `reduce()` instead.
It's important to note that, by default, GlideQuery only fetches the values of the fields you specify with `select()`. If you call `toArray()` without calling `select()` first, the returned objects will only have `sys_id` field.
Also, remember to use `limit()` method when using `toArray()`, to avoid fetching a large number of records and potentially running out of memory.
Resources
- Google Colab for this Post (opens in a new tab)
- Repository for this Post (opens in a new tab)
- Official Langchain tutorial (opens in a new tab)
- StackOverflow question (opens in a new tab)
- GlideQuery Blog Post by ServiceNow (opens in a new tab)
- A GlideQuery Cheat Sheet by Sam Meylan (opens in a new tab)
- Official PySNC documentation (opens in a new tab)