> ## Documentation Index
> Fetch the complete documentation index at: https://notes.kodekloud.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Demo Building Dynamic Context with Custom Data Part 2
> This workflow builds dynamic context from a DataFrame using embeddings and crafts prompts for OpenAI API calls.
In this guide, we'll continue from having a `text` column ready as context. Next, we'll generate word embeddings for each sentence, compute similarities, and dynamically build context for OpenAI API calls—all using a pandas DataFrame.
## 1. Generate Embeddings for the `text` Column
Use your embedding function to vectorize each row in the DataFrame:
```python theme={null}
# Generate embeddings for each row in the 'text' column
df = df.assign(embedding=df["text"].apply(lambda x: text_embedding(x)))
```
Generating embeddings for a large dataset may take some time. For 121 rows, expect multiple API calls.
Inspect the new `embedding` column:
```python theme={null}
df.head()
```
Example output:
```text theme={null}
year_film year_ceremony ceremony category name film winner text embedding
10639 2022 2023 95 actor in a leading role Austin Butler Elvis False Austin Butler got nominated under the category... [–0.0378, –0.0199, …]
10640 2022 2023 95 actor in a leading role Colin Farrell The Banshees of Inisherin False Colin Farrell got nominated under the category... [–0.0082, –0.0100, …]
… … … … … … … … … …
```
To view a specific example:
```python theme={null}
df["text"].iloc[100]
# Output:
# 'Viktor Prášil, Frank Kruse, Markus Stembler, Lars Ginzel and Stefan Korte got nominated under the category, sound, for the film All Quiet on the Western Front but did not win'
```
## 2. Define a Similarity Function
We'll use the dot product to measure vector similarity:
```python theme={null}
import numpy as np
def vector_similarity(vec1, vec2):
return np.dot(np.squeeze(np.array(vec1)), np.squeeze(np.array(vec2)))
```
## 3. Querying the DataFrame with Similarity Search
When asking a question (e.g., “Who won the Best Picture award?”), follow these steps:
1. **Embed the query**
2. **Compute similarity** against each row
3. **Select top candidates**
```python theme={null}
# 1. Embed the query
query = "Who won the Best Picture award?"
query_embedding = text_embedding(query)
# 2. Compute similarity for each row
df["similarity"] = df["embedding"].apply(lambda x: vector_similarity(x, query_embedding))
# 3. Retrieve top 20 matches
top_res = df.nlargest(20, "similarity")
```
## 4. Build the Context Block
Concatenate the top candidates into one string:
```python theme={null}
context = "\n".join(top_res["text"])
print(context)
```
Now craft the prompt and call the API:
````python theme={null}
prompt = f"""
From the data provided in three backticks, respond to the question: {query}
```{context}```
"""
result = get_word_completion(prompt)
print(result)
```text
Expected output:
````
The film "Everything Everywhere All at Once" won the Best Picture award at the 95th Oscar awards.
````text theme={null}
Keep an eye on token limits when concatenating many text entries into `context`.
## 5. Try a Different Question
For example, “Who is the lyricist for RRR?”:
```python
query = "Who is the lyricist for RRR?"
query_embedding = text_embedding(query)
# Recompute similarities and get top matches
df["similarity"] = df["embedding"].apply(lambda x: vector_similarity(x, query_embedding))
top_res = df.nlargest(20, "similarity")
context = "\n".join(top_res["text"])
# Craft prompt and call API
prompt = f"""
From the data provided in three backticks, respond to the question: {query}
```{context}```
"""
result = get_word_completion(prompt)
print(result)
````
Expected response:
```text theme={null}
The lyricist for RRR is Chandra Bose.
```
## 6. Resources and References
| Resource | Description | Link |
| -------------------------------- | -------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------ |
| pandas DataFrame | Data manipulation and analysis | [https://pandas.pydata.org/](https://pandas.pydata.org/) |
| OpenAI Embeddings API | Generate vector embeddings | [https://openai.com/embeddings](https://openai.com/embeddings) |
| NumPy | Numerical computing in Python | [https://numpy.org/](https://numpy.org/) |
| Best Practices for Prompt Design | Tips for crafting effective AI prompts | [https://platform.openai.com/docs/guides/chat/completions-prompt-design](https://platform.openai.com/docs/guides/chat/completions-prompt-design) |
***
This workflow—building dynamic context from a DataFrame using embeddings, then crafting prompts on the fly—unlocks powerful generative AI applications with OpenAI.