Skip to main content

Codebase retrieval

Continue indexes your codebase so that it can later automatically pull in the most relevant context from throughout your workspace. This is done via a combination of embeddings-based retrieval and keyword search. By default, all embeddings are calculated locally with all-MiniLM-L6-v2 and stored locally in ~/.continue/index.

Currently, the codebase retrieval feature is available as the "codebase" and "folder" context providers. You can use them by typing @codebase or @folder in the input box, and then asking a question. The contents of the input box will be compared with the embeddings from the rest of the codebase (or folder) to determine relevant files.

Here are some common use cases where it can be useful:

  • Asking high-level questions about your codebase
    • "How do I add a new endpoint to the server?"
    • "Do we use VS Code's CodeLens feature anywhere?"
    • "Is there any code written already to convert HTML to markdown?"
  • Generate code using existing samples as reference
    • "Generate a new React component with a date picker, using the same patterns as existing components"
    • "Write a draft of a CLI application for this project using Python's argparse"
    • "Implement the foo method in the bar class, following the patterns seen in other subclasses of baz.
  • Use @folder to ask questions about a specific folder, increasing the likelihood of relevant results
    • "What is the main purpose of this folder?"
    • "How do we use VS Code's CodeLens API?"
    • Or any of the above examples, but with @folder instead of @codebase

Here are use cases where it is not useful:

  • When you need the LLM to see literally every file in your codebase
    • "Find everywhere where the foo function is called"
    • "Review our codebase and find any spelling mistakes"
  • Refactoring
    • "Add a new parameter to the bar function and update usages"

Configuration​

There are a few options that let you configure the behavior of the codebase context provider. These can be set in config.json, and are the same for the codebase and folder context providers:

~/.continue/config.json
{
  "contextProviders": [
    {
      "name": "codebase",
      "params": {
        "nRetrieve": 25,
        "nFinal": 5,
        "useReranking": true
      }
    }
  ]
}

nRetrieve​

Number of results to initially retrieve from vector database (default: 25)

nFinal​

Final number of results to use after re-ranking (default: 5)

useReranking​

Whether to use re-ranking, which will allow initial selection of nRetrieve results, then will use an LLM to select the top nFinal results (default: true)

Embeddings providers​

We also support other methods of generating embeddings, which can be configured with the "embeddingsProvider" property in config.json. We currently have built-in support for the following providers:

Transformers.js (currently VS Code only)​

Transformers.js is a JavaScript port of the popular Transformers library. It allows embeddings to be calculated locally in the browser (or in this case inside of the sidebar of your IDE). The model used is all-MiniLM-L6-v2, which is shipped alongside the Continue extension and generates embeddings of size 384.

~/.continue/config.json
{
  "embeddingsProvider": {
    "provider": "transformers.js"
  }
}

Ollama​

Ollama is the easiest way to get up and running with open-source language models. It provides an entirely local REST API for working with LLMs, including generating embeddings. We recommend using an embeddings model like nomic-embed-text:

~/.continue/config.json
{
  "embeddingsProvider": {
    "provider": "ollama",
    "model": "nomic-embed-text",
    "apiBase": "http://localhost:11434" // optional, defaults to http://localhost:11434
  }
}

Voyage AI​

Voyage AI offers the best embeddings for code with their voyage-code-2 model. After obtaining an API key from here, you can configure like this:

~/.continue/config.json
{
  "embeddingsProvider": {
    "provider": "openai",
    "model": "voyage-code-2",
    "apiBase": "https://api.voyageai.com/v1/",
    "apiKey": "<VOYAGE_API_KEY>"
  }
}

OpenAI​

OpenAI's embeddings are high dimensional embeddings that give great performance on both text and code.

Configuration for the text-embedding-3-small model​

This is default. The text-embedding-3-small model offers an outstanding balance between performance and efficiency, suitable for a versatile range of applications.

~/.continue/config.json
{
  "embeddingsProvider": {
    "provider": "openai",
    "model": "text-embedding-3-small",
    "apiBase": "<your custom OpenAI-compatible endpoint>", // optional, defaults to OpenAI's API
    "apiKey": "<OPENAI_API_KEY>"
  }
}

Configuration for the text-embedding-3-large model​

For those requiring the highest level of embedding detail and precision, the text-embedding-3-large model is the better choice.

~/.continue/config.json
{
  "embeddingsProvider": {
    "provider": "openai",
    "model": "text-embedding-3-large",
    "apiBase": "<your custom OpenAI-compatible endpoint>", // optional, defaults to OpenAI's API
    "apiKey": "<OPENAI_API_KEY>"
  }
}

Legacy Model Configuration​

For certain scenarios, you may still find the text-embedding-ada-002 model relevant. Below is the configuration example:

~/.continue/config.json
{
  "embeddingsProvider": {
    "provider": "openai",
    "model": "text-embedding-ada-002",
    "apiBase": "<your custom OpenAI-compatible endpoint>", // optional, defaults to OpenAI's API
    "apiKey": "<OPENAI_API_KEY>"
  }
}

Cohere​

Configuration for the embed-english-v3.0 model. This is the default.

~/.continue/config.json
{
  "embeddingsProvider": {
    "provider": "cohere",
    "model": "embed-english-v3.0",
    "apiKey": "<COHERE_API_KEY>"
  }
}

See Cohere's embeddings for available models. Only embedding models v3 and higher are supported.

Gemini​

Gemini's Text Embedding model is optimized for creating embeddings with 768 dimensions for text of up to 2,048 tokens.

As of May 2024, the only available embedding model from Gemini is text-embedding-004.

~/.continue/config.json
{
  "embeddingsProvider": {
    "provider": "gemini",
    "apiKey": "<GEMINI_API_KEY>"
  }
}

Writing a custom EmbeddingsProvider​

If you have your own API capable of generating embeddings, Continue makes it easy to write a custom EmbeddingsProvider. All you have to do is write a function that converts strings to arrays of numbers, and add this to your config in config.ts. Here's an example:

~/.continue/config.ts
export function modifyConfig(config: Config): Config {
  config.embeddingsProvider = {
    embed: (chunks: string[]) => {
      return Promise.all(
        chunks.map(async (chunk) => {
          const response = await fetch("https://example.com/embeddings", {
            method: "POST",
            body: JSON.stringify({ text: chunk }),
          });
          const data = await response.json();
          return data.embedding;
        }),
      );
    },
  };

  return config;
}

Reranking providers​

The reranker plays a crucial role in refining the results retrieved from your codebase. It processes the initial set of results obtained through embeddings-based retrieval, improving their relevance and accuracy for your queries.

Continue offers several reranking options: cohere, voyage, llm, and free-trial, which can be configured in config.json.

Voyage AI​

Voyage AI offers the best reranking model for code with their rerank-lite-1 model. After obtaining an API key from here, you can configure like this:

~/.continue/config.json
{
  "reranker": {
    "name": "voyage",
    "params": {
      "model": "rerank-lite-1",
      "apiKey": "<VOYAGE_API_KEY>"
    }
  }
}

Cohere​

See Cohere's documentation for rerankers here.

~/.continue/config.json
{
  "reranker": {
    "name": "cohere",
    "params": {
      "model": "rerank-english-v3.0",
      "apiKey": "<COHERE_API_KEY>"
    }
  }
}

LLM​

If you only have access to a single LLM, then you can use it as a reranker. This is discouraged unless truly necessary, because it will be much more expensive and still less accurate than any of the above models trained specifically for the task. Note that this will not work if you are using a local model, for example with Ollama, because too many parallel requests need to be made.

~/.continue/config.json
{
  "reranker": {
    "name": "llm",
    "params": {
      "modelTitle": "My Model Title"
    }
  }
}

The "modelTitle" field must match one of the models in your "models" array in config.json.

Free Trial (Voyage AI)​

Continue offers a free trial of Voyage AI's reranking model.

~/.continue/config.json
{
  "reranker": {
    "name": "free-trial"
  }
}

Customizing which files are indexed​

Continue respects .gitignore files in order to determine which files should not be indexed. If you'd like to exclude additional files, you can add them to a .continueignore file, which follows the exact same rules as .gitignore.

If you want to see exactly what files Continue has indexed, the metadata is stored in ~/.continue/index/index.sqlite. You can use a tool like DB Browser for SQLite to view the tag_catalog table within this file.

If you need to force a refresh of the index, reload the VS Code window with cmd/ctrl + shift + p + "Reload Window".