|
| 1 | +# blindRAG |
| 2 | + |
| 3 | +Retrieval Augmented Generation (RAG) is a technique that grants large language |
| 4 | +models (LLMs) information retrieval capabilities and context that they might be |
| 5 | +missing. Nillion's RAG (blindRAG) uses [SecretLLM](/build/secretLLM/overview), [SecretVault](/build/secret-vault), and the |
| 6 | +[nilQL](/build/nilQL) encryption library. |
| 7 | + |
| 8 | +:::info |
| 9 | +blindRAG lets you to store private information in [SecretVault](/build/secret-vault) and then use it as context when you use the [SecretLLM](/build/secretLLM/overview) chat endpoint. |
| 10 | +::: |
| 11 | + |
| 12 | + |
| 13 | +## Library Overview |
| 14 | +Data owners often possess valuable files that clients wish to query to enhance |
| 15 | +their LLM-based inferences. However, ensuring privacy is a key challenge: data |
| 16 | +owners want to keep their data confidential, and clients are equally concerned |
| 17 | +about safeguarding their queries. blindRAG addresses this challenge by enabling |
| 18 | +secure data sharing and querying. It allows data owners to store their data |
| 19 | +securely in SecretVault while allowing clients to query the data without |
| 20 | +exposing their queries or compromising the data's privacy. The process involves |
| 21 | +leveraging a SecretLLM for secure computation through nilAI. Data owners upload |
| 22 | +their information to SecretVault, while SecretLLM processes client queries and |
| 23 | +retrieves the most relevant results (top-k) without revealing sensitive |
| 24 | +information from either party. |
| 25 | + |
| 26 | +blindRAG supports optional clustering to accelerate query retrieval. Data owners |
| 27 | +locally partition their dataset into clusters, then upload the clusters to |
| 28 | +SecretVault. At query time, SecretLLM first identifies the most relevant cluster |
| 29 | +for the incoming query embedding and then executes RAG within that subset. |
| 30 | +By minimizing the search space, this approach reduces comparison overhead and |
| 31 | +significantly speeds up inference. |
| 32 | + |
| 33 | + |
| 34 | +Let us deep dive into the entities and their roles in the system. |
| 35 | + |
| 36 | +1. **Data Owners:** Securely upload files to SecretVault. Before sending the |
| 37 | + files to SecretVault, they are processed into multiple chunks of data and |
| 38 | + their corresponding embeddings. The embeddings are used for similarity |
| 39 | + search, while the chunks are used to retrieve the actual uploaded files. Once |
| 40 | + the files are encoded into chunks and embeddings, they are blinded before |
| 41 | + being uploaded to SecretVault, where each chunk and embedding is |
| 42 | + secret-shared. Optionally, data owners can locally partition their data |
| 43 | + into clusters and upload the chunks and embeddings along with the |
| 44 | + corresponding cluster information to SecretVault. |
| 45 | + |
| 46 | + For instance, a data owner, wishes to upload the following file to SecretVault and later use it to provide context to SecretLLM: |
| 47 | + :::note Employees Example |
| 48 | + ``` |
| 49 | + Kyle Moore works at Jackson, Gray and Lewis as a Economist. Kyle Moore was born on 1915-09-27 and lives at 6206 Caroline Point, Bishopland, MI 34522. |
| 50 | +
|
| 51 | + Michelle Ross works at Davis-Alvarez as a Tree surgeon. Michelle Ross was born on 1946-09-15 and lives at 33554 Deanna Summit Apt. 813, Hurstshire, IA 55587. |
| 52 | +
|
| 53 | + Danielle Miller works at Bailey and Sons as a Engineer, mining. Danielle Miller was born on 2007-10-22 and lives at 61586 Michael Greens, New Holly, CO 29872. |
| 54 | + ... |
| 55 | + ``` |
| 56 | + ::: |
| 57 | +
|
| 58 | + Let's dive a bit more into the example of employees records. First, data |
| 59 | + owners need to create a schema and a query in SecretVault. If clustering is enabled, |
| 60 | + data owners also create a clusters' schema to store the centroids of |
| 61 | + the clusters. |
| 62 | + <details> |
| 63 | + <summary>Full bootstrap.py</summary> |
| 64 | + ```py reference showGithubLink |
| 65 | + https://github.com/NillionNetwork/blindRAG/blob/main/examples/init/bootstrap.py |
| 66 | + ``` |
| 67 | + </details> |
| 68 | +
|
| 69 | + Now that the schemas and the query are ready, data owners can upload their data. If clustering is enabled, |
| 70 | + data owners start by locally computing the clusters centroids using |
| 71 | + [scikit-learn KMeans](https://scikit-learn.org/stable/modules/generated/sklearn.cluster.KMeans.html) method. |
| 72 | + <details> |
| 73 | + <summary>Full write.py</summary> |
| 74 | + ```py reference showGithubLink |
| 75 | + https://github.com/NillionNetwork/blindRAG/blob/main/examples/data_owner/write.py |
| 76 | + ``` |
| 77 | + </details> |
| 78 | +
|
| 79 | +
|
| 80 | +2. **Client:** The client submits a query to search against the data owners' |
| 81 | + uploaded files in SecretVault, retrieve the most relevant data, and use the |
| 82 | + top-k results for privacy-preserving inference in SecretLLM. Similar to the |
| 83 | + encoding by data owners, the query is processed into its corresponding |
| 84 | + embeddings. If clustering is enabled, the most relevant cluster is first |
| 85 | + identified and RAG is executed over this cluster. |
| 86 | +
|
| 87 | + Going back to our example, the client can query SecretLLM asking about Michelle: |
| 88 | + :::note Employees Example |
| 89 | + ``` |
| 90 | + Who is Michelle Ross? |
| 91 | + ``` |
| 92 | + ::: |
| 93 | +
|
| 94 | + Here is an example of how clients can run such a query: |
| 95 | + <details> |
| 96 | + <summary>Full query.py</summary> |
| 97 | + ```py reference showGithubLink |
| 98 | + https://github.com/NillionNetwork/blindRAG/blob/main/examples/client/query.py |
| 99 | + ``` |
| 100 | + </details> |
| 101 | +
|
| 102 | +
|
| 103 | +3. **SecretVault:** SecretVault stores the blinded chunks and embeddings |
| 104 | + provided by data owners. When a client submits a query, SecretVault computes |
| 105 | + the differences between the query's embeddings and each stored embedding in a |
| 106 | + privacy-preserving manner. If clustering is enabled, SecretVault also stores the |
| 107 | + cluster centroids in a separate schema. In the original schema, the blinded chunks |
| 108 | + and embeddings are stored along with the corresponding centroid. |
| 109 | +
|
| 110 | +
|
| 111 | +4. **SecretLLM:** SecretLLM connects to SecretVault to fetch the blinded |
| 112 | + differences between the query and the stored embeddings and then compute the |
| 113 | + closest matches. If clustering is enabled, SecretLLM starts by retrieving the |
| 114 | + centroid points. Finally, it uses the top k matches for inference. |
| 115 | +
|
| 116 | + Lastly, the client can query SecretLLM asking about Michelle: |
| 117 | + :::note Employees Example |
| 118 | + ``` |
| 119 | + Michelle Ross is an engineer who works at Bailey and Sons, specializing in mining. She was born on October 22, 2007, and lives at 61586 Michael Greens, New Holly, CO 29872. |
| 120 | + ``` |
| 121 | + ::: |
| 122 | +
|
| 123 | +
|
| 124 | +You can reproduce the example above by following the [README](https://github.com/NillionNetwork/blindrag). |
| 125 | +
|
| 126 | +## Implementation |
| 127 | +
|
| 128 | +blindRAG is a standalone library available through |
| 129 | +[PyPI](https://pypi.org/project/blindrag) and open-source on |
| 130 | +[GitHub](https://github.com/NillionNetwork/blindrag). Developers can use blindRAG as |
| 131 | +a feature of [SecretLLM](https://docs.nillion.com/build/secretLLM/quickstart) to |
| 132 | +enhance the inference with context that has been uploaded to [SecretVault](https://docs.nillion.com/build/secret-vault). |
| 133 | +
|
| 134 | +
|
| 135 | +### Performance Expectations |
| 136 | +
|
| 137 | +We have performed a series of benchmarks to evaluate the performance of blindRAG with and without clustering. |
| 138 | +Currently, blindRAG scales linearly to the number of rows stored in SecretVault. |
| 139 | +The following table shows latency to upload to SecretVault multiple paragraphs of a few sentences long, as well as the runtime for AI inference using SecretLLM with blindRAG. |
| 140 | +
|
| 141 | +<table> |
| 142 | + <thead> |
| 143 | + <tr> |
| 144 | + <th rowspan="2">Number of Paragraphs Stored in SecretVault</th> |
| 145 | + <th colspan="2">RAG Time (sec.)</th> |
| 146 | + <th colspan="2">Query Time (Inference + RAG, sec.)</th> |
| 147 | + </tr> |
| 148 | + <tr> |
| 149 | + <th>No Clusters</th> |
| 150 | + <th>5 Clusters</th> |
| 151 | + <th>No Clusters</th> |
| 152 | + <th>5 Clusters</th> |
| 153 | + </tr> |
| 154 | + </thead> |
| 155 | + <tbody> |
| 156 | + <tr> |
| 157 | + <td>1</td> |
| 158 | + <td>0.2</td> |
| 159 | + <td> - </td> |
| 160 | + <td>2.4</td> |
| 161 | + <td> - </td> |
| 162 | + </tr> |
| 163 | + <tr> |
| 164 | + <td>10</td> |
| 165 | + <td>0.4</td> |
| 166 | + <td> - </td> |
| 167 | + <td>3.1</td> |
| 168 | + <td> - </td> |
| 169 | + </tr> |
| 170 | + <tr> |
| 171 | + <td>100</td> |
| 172 | + <td>2.3 </td> |
| 173 | + <td> 1.7 </td> |
| 174 | + <td>2.9</td> |
| 175 | + <td> 2.1 </td> |
| 176 | + </tr> |
| 177 | + <tr> |
| 178 | + <td>1 000</td> |
| 179 | + <td>5.8</td> |
| 180 | + <td>2.5</td> |
| 181 | + <td>7.0</td> |
| 182 | + <td>3.2</td> |
| 183 | + </tr> |
| 184 | + <tr> |
| 185 | + <td>5 000</td> |
| 186 | + <td>20.0</td> |
| 187 | + <td>5.7</td> |
| 188 | + <td>25.1</td> |
| 189 | + <td>5.9</td> |
| 190 | + </tr> |
| 191 | + <tr> |
| 192 | + <td>10 000</td> |
| 193 | + <td>39.2</td> |
| 194 | + <td>10.0</td> |
| 195 | + <td>47.5</td> |
| 196 | + <td>8.9</td> |
| 197 | + </tr> |
| 198 | + <tr> |
| 199 | + <td>20 000</td> |
| 200 | + <td>74.7</td> |
| 201 | + <td>11.3</td> |
| 202 | + <td>92.5</td> |
| 203 | + <td>19.8</td> |
| 204 | + </tr> |
| 205 | + </tbody> |
| 206 | +</table> |
| 207 | +
|
| 208 | +Additionally, using multiple concurrent users, the query time for inference with blindRAG increases. |
| 209 | +Performing inference with blindRAG with a content of 100 paragraphs takes approximately 5 seconds for a single user, while with ten concurrent users the inference time for the same content goes up to almost 9 seconds. |
| 210 | +We're developing new research to further accelerate blindRAG and make it more scalable, stay tuned! |
0 commit comments