An application programming interface (API) lets a program send a request to another service and receive a response. Think of it as a service counter: your Python code submits a prompt to a generative AI model, then uses the returned text in an analysis. Unlike a web chat, API access can fit into a repeatable workflow or a custom analytics application. section 1.3.3
A practical model-API workflow has a few checkpoints:
- Set up the provider’s library and API key, then follow its current documentation. SDKs can reduce routine coding, but check their compatibility and support. section 1.3.3
- Send a prompt and any relevant model parameters; collect the response in Python rather than assuming it will arrive in exactly the format you want. section 5.5.3
- Inspect results and handle failures. A model may generate code for an outdated API interface, and even a successful request may return text that needs careful parsing. section 5.5.3
From a customer question to review analysis
Suppose a team wants to find positive Portuguese customer reviews. An application could request candidate search keywords through an API, then run its own code to find matching reviews. That is feasible, but the returned keyword list needs inspection before it drives a search: consistent formatting is not guaranteed. section 5.5.3
Listing 6.1 Sentiment analysis with ChatGPT's intrinsic language capabilities
from openai import OpenAI
# Replace 'your_openai_api_key' with your actual OpenAI API key
client = OpenAI(
api_key= "your-api-key",
)
def get_sentiment(review):
response = client.chat.completions.create(
messages=[
{
"role": "user",
"content": f"The sentiment of this review is: {review}",
}
],
model="gpt-4-0125-preview",
)
completion = response.choices[0].message.content
if "positive" in completion:
return "positive"
elif "neutral" in completion:
return "neutral"
elif "negative" in completion:
return "negative"
else:
return "unknown"
# Analyze the reviews and store the output (manually adapted)
sentiments = []
for review in reviews:
sentiments.append(get_sentiment(review))
df["GPT4"] = sentiments
The Python listing sends each review to ChatGPT and extracts a sentiment label from the response. Notice that its simple word-matching rules can miss nuanced or mixed answers. section 6.2.3
Scale does not remove the need to validate. In a 500-review comparison using review scores as the reference, the API-based method reached 0.74 sensitivity and 0.93 specificity for positive reviews; basic keyword search reached 0.45 and 0.90. APIs also do not give a generative AI automatic access to your databases or other data sources—connecting those systems requires additional integration. section 6.2.3 section 1.2.3
Table 6.2 Quality of sentiment analysis with ChatGPT 4 API
|
Sentiment analysis (
n=500)
|
Number of reviews classified as positive
|
Sensitivity
|
Specificity
|
|---|---|---|---|
|
Reference review score
|
320 (64%)
true positives
|
1
|
1
|
|
Basic keyword search
|
163 (33%)
|
0.45
|
0.9
|
|
ChatGPT 4 API
|
250 (50%)
|
0.74
|
0.93
|
Compare the API method with keyword search against the same review-score reference, rather than judging it by plausible-looking individual responses. section 6.2.3
