198 lines
6.5 KiB
Markdown
198 lines
6.5 KiB
Markdown
# [Sem_Cache](https://github.com/kjannette/semantic-cache-script)
|
||
|
||
Sem_Cache is a tiny command-line tool written in [Python](https://www.python.org/) that
|
||
demonstrates semantic caching for Large Language Model (LLM) queries. It uses the
|
||
[Sentence Transformers](https://www.sbert.net/) library with [NumPy](https://numpy.org/)
|
||
to compute vector embeddings and [cosine similarity](https://en.wikipedia.org/wiki/Cosine_similarity)
|
||
between text inputs, determining whether an incoming query is similar enough to a
|
||
cached query to return a stored response instead of making a new LLM Application
|
||
Programming Interface (API) call.
|
||
|
||
Released under the
|
||
[GPL Version 3](https://opensource.org/license/gpl-3-0). Initial release: August 2026.
|
||
|
||
This project is intended for developers building LLM applications who want
|
||
to reduce API costs, latency, and redundant computations by caching semantically
|
||
equivalent queries.
|
||
|
||
It can be used to quickly verify the reliability of a semantic caching layer by generating a cosine score using two test sentences, input as command line strings.
|
||
|
||
The output can be diff’d against the behavior of the system under development for an instant “sanity check.”
|
||
|
||
---
|
||
|
||
## How It Works
|
||
|
||
If two queries mean the same thing, they should return the same answer. Rather than comparing strings character-by-character, it converts each query into a dense vector embedding using the [all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) model (a lightweight transformer that runs locally). It then calculates the cosine similarity between the cached query embedding and the new query embedding.
|
||
|
||
If the similarity score meets or exceeds the threshold (default: 0.92), the system
|
||
declares a **cache hit** and would return the cached response. Otherwise, it declares
|
||
a **cache miss** and routes the request to the LLM.
|
||
|
||
---
|
||
|
||
## Getting Started
|
||
|
||
### Prerequisites
|
||
|
||
- [Python](https://www.python.org/) 3.10 or higher
|
||
- [pip](https://pip.pypa.io/) (Python package installer)
|
||
|
||
### Installation
|
||
|
||
Clone the repository and set up a virtual environment:
|
||
|
||
```bash
|
||
git clone https://github.com/kjannette/semantic-cache-script.git
|
||
cd semantic-cache-script
|
||
python -m venv .venv
|
||
source .venv/bin/activate # On Windows: .venv\Scripts\activate
|
||
```
|
||
|
||
Install the required dependencies:
|
||
|
||
```bash
|
||
pip install numpy sentence-transformers
|
||
```
|
||
|
||
The first run will download the all-MiniLM-L6-v2 model (approximately 90 MB).
|
||
|
||
### Usage
|
||
|
||
Run the script from the command line:
|
||
|
||
```bash
|
||
python sem_cache.py
|
||
```
|
||
|
||
The program will prompt you to enter two sentences:
|
||
|
||
1. **Sentence 1**: The baseline cached query (simulating a query already in the cache)
|
||
2. **Sentence 2**: The new incoming query (simulating a user's new request)
|
||
|
||
### Example Session
|
||
|
||
```
|
||
============================================================
|
||
LOCAL SEMANTIC CACHE SIMULATOR
|
||
============================================================
|
||
Loading local embedding model (all-MiniLM-L6-v2)...
|
||
|
||
Model loaded successfully.
|
||
------------------------------------------------------------
|
||
Enter Sentence 1 (The Baseline Cached Query):
|
||
> What is the weather like today?
|
||
|
||
Enter Sentence 2 (The New Incoming Query):
|
||
> How's the weather today?
|
||
|
||
Generating local embeddings and performing vector math...
|
||
------------------------------------------------------------
|
||
RESULTS:
|
||
-> Calculated Cosine Similarity: 0.9412
|
||
-> Target Safety Threshold: 0.9200
|
||
|
||
[CACHE HIT] Returning cached response.
|
||
============================================================
|
||
```
|
||
|
||
In this example, the two queries are semantically equivalent, so the system returns
|
||
a cache hit with a similarity score of 0.9412.
|
||
|
||
---
|
||
|
||
## Configuration
|
||
|
||
The similarity threshold is set to 0.92 by default. This is a highly-conservative value, set to reduce false positives (treating dissimilar queries as matches). To adjust the threshold,
|
||
modify the `THRESHOLD` constant in `sem_cache.py`:
|
||
|
||
```python
|
||
THRESHOLD = 0.92
|
||
```
|
||
|
||
Lower values increase cache hit rates but risk returning incorrect cached responses.
|
||
Higher values reduce false positives but may miss valid semantic matches.
|
||
|
||
---
|
||
|
||
## Project Structure
|
||
|
||
```
|
||
semantic-cache-script/
|
||
├── sem_cache.py # Main application script
|
||
├── README.md # This file
|
||
├── .gitignore # Git ignore patterns
|
||
└── .venv/ # Python virtual environment (not tracked)
|
||
```
|
||
|
||
---
|
||
|
||
## Participation
|
||
|
||
### Bug Reports
|
||
|
||
Bug reports are accepted via [Git issues](https://github.com/kjannette/semantic-cache-script/issues).
|
||
Include the Python version, operating system, input sentences, and full error output
|
||
when reporting issues.
|
||
|
||
### Pull Requests
|
||
|
||
Pull requests are accepted for review. Project author makes no guarantee that
|
||
contributions will be merged. No Contributor License Agreement (CLA) is required.
|
||
|
||
### Code Style
|
||
|
||
This project follows [PEP 8](https://peps.python.org/pep-0008/) style guidelines.
|
||
|
||
---
|
||
|
||
## Author
|
||
|
||
- @ sjDev
|
||
|
||
---
|
||
|
||
## Ideology
|
||
|
||
This project does not have a formal Code of Conduct.
|
||
|
||
Sem_Cache is a standalone, free software project. It is not associated with a
|
||
for-profit company or "open core" offering. The software runs locally and does
|
||
not transmit activity data off the device where it runs (the embedding model
|
||
executes on-device; no external API calls are made by the caching logic itself).
|
||
|
||
---
|
||
|
||
## Roadmap / TO-DO
|
||
|
||
(This README also serves as a development notebook.)
|
||
|
||
### Planned Features
|
||
|
||
- [ ] Print result to command line (completed in current version)
|
||
- [ ] Perform and output quantified metrics of what a cache hit conserves
|
||
|
||
### Possible Metrics to Implement
|
||
|
||
1. **Estimated completion tokens** — Calculate tokens saved by popular model
|
||
(assuming cache did not exist or cache misses)
|
||
2. **Latency** — Measure the API call latency to the LLM that would be avoided
|
||
on cache hit
|
||
3. **Rate limit / quota impact** — Theoretical calculation of quota preservation
|
||
(low priority)
|
||
4. **Cost of redundant embedding** — Quantify savings from not re-embedding
|
||
identical strings across runs, retries, and re-indexes
|
||
5. **Provider outage resilience** — Document how cached responses provide
|
||
continuity when upstream services are unavailable
|
||
6. **Reduction of nondeterminism** — Address unreliable tests and unreproducible
|
||
bugs by enabling diff comparisons between runs when model output varies
|
||
7. **Concurrency waste avoidance** — Quantify savings when N users submit
|
||
identical queries (e.g., 50 users click the same button, ask the same
|
||
question; without cache, cost = 50x)
|
||
|
||
---
|
||
|
||
## License
|
||
|
||
This softeare is released under the [GNU General Public License Version 3](https://opensource.org/license/gpl-3-0).
|