Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add context #9

Open
wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Add context #9

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
a48e7e5
Add context
Nov 26, 2023
3cddae7
Explain context in README
j-simian Nov 27, 2023
a3a3d32
Merge branch 'main' into main
j-simian Nov 27, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
89 changes: 40 additions & 49 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,14 +1,13 @@
# VectorDB

[![](https://dcbadge.vercel.app/api/server/aDNg6E9szy?compact=true&style=flat)](https://discord.gg/aDNg6E9szy) [![Twitter](https://img.shields.io/twitter/follow/KagiHQ?style=social)](https://twitter.com/KagiHQ) [![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](https://opensource.org/license/mit/)
[![](https://dcbadge.vercel.app/api/server/aDNg6E9szy?compact=true&style=flat)](https://discord.gg/aDNg6E9szy) [![Twitter](https://img.shields.io/twitter/follow/KagiHQ?style=social)](https://twitter.com/KagiHQ) [![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](https://opensource.org/license/mit/)

VectorDB is a simple, lightweight, fully local, end-to-end solution for using embeddings-based text retrieval.

Thanks to its low latency and small memory footprint, VectorDB is used to power AI features inside [Kagi Search](https://kagi.com).

Check an [example Colab notebook](https://colab.research.google.com/drive/1pecKGCCru_Jvx7v0WRNrW441EBlcS5qS#scrollTo=Eh6o8m7d8eOk) where this is used to filter the content of [Kagi Small Web](https://kagi.com/smallweb) RSS feed based on stated user interests.


## Installation

To install VectorDB, use pip:
Expand All @@ -19,7 +18,7 @@ pip install vectordb2

## Usage

Quick example that loads data into memory, and runs retrieval. All data will be handled locally, including embeddings and vector search, completely trasparent for the user with maximum possible performance.
Quick example that loads data into memory, and runs retrieval. All data will be handled locally, including embeddings and vector search, completely trasparent for the user with maximum possible performance.

```python
from vectordb import Memory
Expand Down Expand Up @@ -53,55 +52,52 @@ This returns the chunks with the added metadata and the vector distance (where 0

## Options


**Memory(memory_file=None, chunking_strategy={"mode":"sliding_window"},
embeddings="normal")**

- `memory_file`: _Optional._ Path to the memory file. If provided, memory will persist to disk and loaded/saved to this file.
- `chunking_strategy`: _Optional._ Dictionary containing the chunking mode.

- `memory_file`: *Optional.* Path to the memory file. If provided, memory will persist to disk and loaded/saved to this file.
- `chunking_strategy`: *Optional.* Dictionary containing the chunking mode.

Options:\
`{'mode':'sliding_window', 'window_size': 240, 'overlap': 8}` (default)\
Options:\
`{'mode':'sliding_window', 'window_size': 240, 'overlap': 8}` (default)\
`{'mode':'paragraph'}`
- `embeddings`: *Optional.*

Options:\

- `embeddings`: _Optional._

Options:\
`fast` - Uses Universal Sentence Encoder 4\
`normal` - Uses "BAAI/bge-small-en-v1.5" (default)\
`best` - Uses "BAAI/bge-base-en-v1.5"\
`multilingual` - Uses Universal Sentence Encoder Multilingual Large 3


You can also specify a custom HuggingFace model by name eg. `TaylorAI/bge-micro-v2`. See also [Pretrained models](https://www.sbert.net/docs/pretrained_models.html) and [MTEB](https://huggingface.co/spaces/mteb/leaderboard).
You can also specify a custom HuggingFace model by name eg. `TaylorAI/bge-micro-v2`. See also [Pretrained models](https://www.sbert.net/docs/pretrained_models.html) and [MTEB](https://huggingface.co/spaces/mteb/leaderboard).

**Memory.save(texts, metadata, memory_file=None)**

Save content to memory. Metadata will be automatically optimized to use less resources.

- `texts`: *Required.* Text or list of texts to be saved.
- `metdata`: *Optional.* Metadata or list of metadata associated with the texts.
- `memory_file`: *Optional.* Path to persist the memory file. By default
- `texts`: _Required._ Text or list of texts to be saved.
- `metdata`: _Optional._ Metadata or list of metadata associated with the texts.
- `memory_file`: _Optional._ Path to persist the memory file. By default

**Memory.search(query, top_n=5, unique=False, batch_results="flatten")**

Search inside memory.

- `query`: *Required.* Query text or list of queries (see `batch_results` option below for handling results for a list).
- `top_n`: *Optional.* Number of most similar chunks to return (default: 5).
- `unique`: *Optional.* Return only items chunks from unique original texts (additional chunks coming from the same text will be ignored). Note this may return less chhunks than requested (default: False).
- `batch_results`: *Optional.* When input is a list of queries, output algorithm can be "flatten" or "diverse". Flatten returns true nearest neighbours across all input queries, meaning all results could come from just one query. "diverse" attempts to spread out the results, so that each query's nearest neighbours are equally added (neareast first across all queries, than 2nd nearest and so on). (default: "flatten")
- `query`: _Required._ Query text or list of queries (see `batch_results` option below for handling results for a list).
- `top_n`: _Optional._ Number of most similar chunks to return (default: 5).
- `unique`: _Optional._ Return only items chunks from unique original texts (additional chunks coming from the same text will be ignored). Note this may return less chunks than requested (default: False).
- `batch_results`: _Optional._ When input is a list of queries, output algorithm can be "flatten" or "diverse". Flatten returns true nearest neighbours across all input queries, meaning all results could come from just one query. "diverse" attempts to spread out the results, so that each query's nearest neighbours are equally added (neareast first across all queries, than 2nd nearest and so on). (default: "flatten")
- `context` Optional. When `context = n` (default 0), n chunks before and after the result in the original text are returned along with the result. This is useful when using sliding window with smaller chunks, or paragraph chunking when you have relatively short paragraphs.

**Memory.clear()**

Clears the memory.


**Memory.dump()**

Prints the contents of the memory.


## Example

```python
Expand Down Expand Up @@ -184,6 +180,7 @@ print(results)
```

Output:

```json
[
{
Expand All @@ -208,40 +205,34 @@ Output:

## Embeddings performance analysis


We constantly evaluate embedding models using standardized benchmarks (higher is better). Average latency is measured locally on CPU (lower is better). Benchmark data pulled from [MTEB](https://huggingface.co/spaces/mteb/leaderboard).



| Model | Latency | Benchmark 1 | Benchmark 2 | Benchmark 3 | Benchmark 4 |
|-----------------------------------------------|----------|-------------|-------------|-------------|-------------|
| all-mpnet-base-v2 | 6.12 s | 80.28 | 65.07 | 43.69 | 83.04 |
| all-MiniLM-L6-v2 | 1.14 s | 78.9 | 63.05 | 42.35 | 82.37 |
| BAAI/bge-large-en-v1.5 | 20.8 s | 83.11 | 75.97 | 46.08 | 87.12 |
| BAAI/bge-base-en-v1.5 | 6.48 s | 82.4 | 75.53 | 45.77 | 86.55 |
| BAAI/bge-small-en-v1.5 | 1.85 s | 81.59 | 74.14 | 43.82 | 84.92 |
| TaylorAI/bge-micro-v2 | 0.671 s | 78.65 | 68.04 | 39.18 | 82.81 |
| TaylorAI/gte-tiny | 1.25 s | 80.46 | 70.35 | 42.09 | 82.83 |
| thenlper/gte-base | 6.28 s | 82.3 | 73.01 | 46.2 | 84.57 |
| thenlper/gte-small | 2.14 s | 82.07 | 72.31 | 44.89 | 83.54 |
| universal-sentence-encoder-large/5 | 0.769 s | 74.05 | 67.9 | 37.82 | 79.53 |
| universal-sentence-encoder-multilingual-large/3| 1.02 s | 75.35 | 65.78 | 35.06 | 79.62 |
| universal-sentence-encoder-multilingual/3 | 0.162 s | 75.39 | 63.42 | 34.82 | 75.43 |
| universal-sentence-encoder/4 | 0.019 s | 72.04 | 64.45 | 35.71 | 76.23 |

*Relative embeddings latency on CPU*
We constantly evaluate embedding models using standardized benchmarks (higher is better). Average latency is measured locally on CPU (lower is better). Benchmark data pulled from [MTEB](https://huggingface.co/spaces/mteb/leaderboard).

| Model | Latency | Benchmark 1 | Benchmark 2 | Benchmark 3 | Benchmark 4 |
| ----------------------------------------------- | ------- | ----------- | ----------- | ----------- | ----------- |
| all-mpnet-base-v2 | 6.12 s | 80.28 | 65.07 | 43.69 | 83.04 |
| all-MiniLM-L6-v2 | 1.14 s | 78.9 | 63.05 | 42.35 | 82.37 |
| BAAI/bge-large-en-v1.5 | 20.8 s | 83.11 | 75.97 | 46.08 | 87.12 |
| BAAI/bge-base-en-v1.5 | 6.48 s | 82.4 | 75.53 | 45.77 | 86.55 |
| BAAI/bge-small-en-v1.5 | 1.85 s | 81.59 | 74.14 | 43.82 | 84.92 |
| TaylorAI/bge-micro-v2 | 0.671 s | 78.65 | 68.04 | 39.18 | 82.81 |
| TaylorAI/gte-tiny | 1.25 s | 80.46 | 70.35 | 42.09 | 82.83 |
| thenlper/gte-base | 6.28 s | 82.3 | 73.01 | 46.2 | 84.57 |
| thenlper/gte-small | 2.14 s | 82.07 | 72.31 | 44.89 | 83.54 |
| universal-sentence-encoder-large/5 | 0.769 s | 74.05 | 67.9 | 37.82 | 79.53 |
| universal-sentence-encoder-multilingual-large/3 | 1.02 s | 75.35 | 65.78 | 35.06 | 79.62 |
| universal-sentence-encoder-multilingual/3 | 0.162 s | 75.39 | 63.42 | 34.82 | 75.43 |
| universal-sentence-encoder/4 | 0.019 s | 72.04 | 64.45 | 35.71 | 76.23 |

_Relative embeddings latency on CPU_
![Embeddings Latency on CPU](images/speed_cpu.png)

*Relative embeddings latency on GPU*
_Relative embeddings latency on GPU_
![Embeddings Latency on GPU](images/speed_gpu.png)


![Embeddings Quality](images/quality.png)

![Scatter of Embeddings](images/scatter.png)



## Vector search performance analysis

VectorDB is also optimized for speed of retrieval. We automatically uses [Faiss](https://github.com/facebookresearch/faiss) for low number of chunks (<4000) and [mrpt](https://github.com/vioshyvo/mrpt) for high number of chunks to ensure maximum performance across the spectrum of use cases.
Expand Down
23 changes: 14 additions & 9 deletions vectordb/memory.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -133,7 +133,7 @@ def save(
Storage(self.memory_file).save_to_disk([{"memory": self.memory, "metadata" :self.metadata_memory}])

def search(
self, query: str, top_n: int = 5, unique: bool = False, batch_results: str = "flatten"
self, query: str, top_n: int = 5, unique: bool = False, batch_results: str = "flatten", context: int = 0
) -> List[Dict[str, Any]]:
"""
Searches for the most similar chunks to the given query in memory.
Expand Down Expand Up @@ -170,14 +170,19 @@ def search(
) # Use seen_text_indices instead of seen_meta_indices
indices = unique_indices

results = [
{
"chunk": self.memory[i[0]]["chunk"],
"metadata": self.metadata_memory[self.memory[i[0]]["metadata_index"]],
"distance": i[1],
}
for i in indices
]
results = []
for i in indices:
chunk_with_context = ""
for j in range(-context, context+1):
chunk_with_context += self.memory[i[0]+j]["chunk"] if i[0]+j >= 0 and i[0]+j < len(self.memory) else ""
results += [
{
"chunk": chunk_with_context,
"metadata": self.metadata_memory[self.memory[i[0]]["metadata_index"]],
"distance": i[1],
}
]


return results

Expand Down