Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/pramsey/pgsql-openai
Helper functions for working with the OpenAI API
https://github.com/pramsey/pgsql-openai
Last synced: 14 days ago
JSON representation
Helper functions for working with the OpenAI API
- Host: GitHub
- URL: https://github.com/pramsey/pgsql-openai
- Owner: pramsey
- Created: 2024-11-04T23:32:10.000Z (3 months ago)
- Default Branch: main
- Last Pushed: 2024-11-05T18:13:20.000Z (2 months ago)
- Last Synced: 2024-11-05T18:38:38.279Z (2 months ago)
- Language: PLpgSQL
- Size: 49.8 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
- awesome-repositories - pramsey/pgsql-openai - Helper functions for working with the OpenAI API (PLpgSQL)
README
# OpenAI API Client for PostgreSQL
Using the [http extension](https://github.com/pramsey/pgsql-http) for web access, this extension provides utility functions to make it easier to work against the OpenAI API. The API is avilable from [OpenAI](https://platform.openai.com/docs/overview) of course, but also from other tools like the [Ollama](https://ollama.com) runner for local AI models.
## Uses
The most obvious uses for AI in the database are:
- **Classification and text analysis**. Tie an AI trigger to freeform text fields in your database, so that new data (like reviews and comments) are automatically analyzed and categorized (sentiment analysis for example).
- **Similarity searching.** In conjunction with the [pgvector](https://github.com/pgvector/pgvector) extension, find the "N most similar" items in a corpus. Recommendation engines, for example, can find similar items in a corpus to the one a user is currently looking at.
- **Retrieval-augmented generation.** A combination of basic chatbot technology and similarity searching to improve the results of a natural language AI query by feeding the model with relevant contextual documents before posing a query.
## Setup
### Using OpenAI
* Set up an [OpenAI API key](https://help.openai.com/en/articles/4936850-where-do-i-find-my-openai-api-key)
* Create a working database
* Enable the HTTP extension
```sql
CREATE EXTENSION http;
```
* Set the up the session keys
```sql
SET openai.api_key = 'your_api_key_here';
SET openai.api_uri = 'https://api.openai.com/v1/';
SET openai.prompt_model = 'gpt-4o-mini';
SET openai.embedding_model = 'text-embedding-3-small';
```
* (Optionally) make the keys persistent
```sql
ALTER DATABASE your_db SET openai.api_key = 'your_api_key_here';
ALTER DATABASE your_db SET openai.api_uri = 'https://api.openai.com/v1/';
ALTER DATABASE your_db SET openai.prompt_model = 'gpt-4o-mini';
ALTER DATABASE your_db SET openai.embedding_model = 'text-embedding-3-small';
```
### Using Ollama
If you have a workstation with a reasonable amount of memory (16GB or more) you can consider running a mid-sized model locally instead of using OpenAI services.
* [Download Ollama](https://ollama.com) for your workstation
* Verify you can run `ollama`
* then `ollama pull llama3.1:latest`
* and `ollama pull mxbai-embed-large`
* Set the up the session keys
```sql
SET openai.api_uri = 'http://127.0.0.1:11434/v1/';
SET openai.api_key = 'none';
SET openai.prompt_model = 'llama3.1:latest';
SET openai.embedding_model = 'mxbai-embed-large';
```
* (Optionally) make the keys persistent
```sql
ALTER DATABASE your_db SET openai.api_uri = 'http://127.0.0.1:11434/v1/';
ALTER DATABASE your_db SET openai.api_key = 'none';
ALTER DATABASE your_db SET openai.prompt_model = 'llama3.1:latest';
ALTER DATABASE your_db SET openai.embedding_model = 'mxbai-embed-large';
```
## Usage
### Functions
- `openai.models()` returns `setof record`. Returns a listing of the AI models available via the API.
- `openai.prompt(context text, query text, [model text])` returns `text`. Set up context and then pass in a starting prompt for the model.
- `openai.vector(query text, [model text])` returns `text`. For use with embedding models ([Ollama](https://ollama.com/blog/embedding-models), [OpenAI](https://platform.openai.com/docs/guides/embeddings#embedding-models)) only, returns a JSON-formatted float array suitable for direct input into pgvector. Using non-embedding models to generate embeddings will result in extremely poor search results.
### Basic Examples
Read back the settings your database is using.
```sql
SHOW openai.api_uri;
```
```
openai.api_uri
----------------------------
http://127.0.0.1:11434/v1/
```
List all the models running behind the API you are connecting to. If it is a local API like [Ollama](https://ollama.com/), probably a short list. If it is a hosted API like OpenAI, probably a long one.
```sql
SELECT openai.models();
```
```
SELECT * FROM openai.models();
id | object | created | owned_by
--------------------------+--------+---------------------+----------
mxbai-embed-large:latest | model | 2024-11-04 20:48:39 | library
llama3.1:latest | model | 2024-07-25 22:45:02 | library
```
Query the model about factual matters. For small models or obscure facts, this may likely return hallucinated results.
```sql
SELECT openai.prompt(
'You are a very smart physics teacher.',
'What is the speed of light?'
);
```
```
A classic question! The speed of light, in a vacuum (a completely
empty space with no air or matter), is approximately...
(dramatic pause) ...299,792,458 meters per second! +
```
Use the model to summarize a free-form text input. In the extreme, a sentiment analysis prompt can instruct the model to use a fixed set of output tokens, for nice programmatic filtering of free-form inputs.
```sql
SELECT openai.prompt(
'You are an advanced sentiment analysis model. Read the given
feedback text carefully and classify it as one of the
following sentiments only: "positive", "neutral", or
"negative". Respond with exactly one of these words
and no others, using lowercase and no punctuation',
'I enjoyed the setting and the service and the bisque was
great.' );
```
```
prompt
----------
positive
```
Calculate the embedding vector for a given piece of input. Useful for "find the most similar" queries or as an input to [retrieval augmented generation](https://en.wikipedia.org/wiki/Retrieval-augmented_generation) (RAG) systems.
```sql
SELECT openai.vector('A lovely orange pumpkin pie recipe.');
```
```
[-0.021690376, 0.003092674, -0.00023757249, 0.0059805233,
0.0024175171, 0.013349159, 0.006348481, 0.016774092,
0.0051014116, 0.026803626, 0.015113969, 0.0031985058,
...
-0.020051343, 0.006571748, 0.008234819, 0.010086719,
-0.0071006618, -0.020877795, -0.022467814, 0.010012546,
0.0008801813, -0.0006236545, 0.016922941, -0.011781357]
```
## Debugging
To see what's going on inside the functions, set the debug level:
```sql
SET client_min_messages = debug;
```
To work around HTTP timeouts, increase the timeout interval:
```sql
SELECT http_set_curlopt('CURLOPT_TIMEOUT', '10');
```