# Convex RAG Component

[![npm version](https://badge.fury.io/js/@convex-dev%2Frag.svg)](https://badge.fury.io/js/@convex-dev%2Frag)

<!-- START: Include on https://convex.dev/components -->

A component for semantic search, usually used to look up context for LLMs. Use
with an Agent for Retrieval-Augmented Generation (RAG).

[![Use AI to search HUGE amounts of text with the RAG Component](https://thumbs.video-to-markdown.com/1ff18153.jpg)](https://youtu.be/dGmtAmdAaFs)

## ✨ Key Features

- **Add Content**: Add or replace content with text chunks and embeddings.
- **Semantic Search**: Vector-based search using configurable embedding models
- **Namespaces**: Organize content into namespaces for per-user search.
- **Custom Filtering**: Filter content with custom indexed fields.
- **Importance Weighting**: Weight content by providing a 0 to 1 "importance".
- **Chunk Context**: Get surrounding chunks for better context.
- **Graceful Migrations**: Migrate content or whole namespaces without
  disruption.

Found a bug? Feature request?
[File it here](https://github.com/get-convex/rag/issues).

## Installation

```sh
npm install @convex-dev/rag ai
```

The [AI SDK](https://ai-sdk.dev) (`ai@^7`) is a peer dependency, so your app and
the component always share one copy.

Create a `convex.config.ts` file in your app's `convex/` folder and install the
component by calling `use`:

```ts
// convex/convex.config.ts
import { defineApp } from "convex/server";
import rag from "@convex-dev/rag/convex.config.js";

const app = defineApp();
app.use(rag);

export default app;
```

Run `npx convex codegen` if `npx convex dev` isn't already running.

## Basic Setup

```ts
// convex/example.ts
import { components } from "./_generated/api";
import { RAG } from "@convex-dev/rag";
// Any AI SDK model that supports embeddings will work.
import { openai } from "@ai-sdk/openai";

const rag = new RAG(components.rag, {
  textEmbeddingModel: openai.embedding("text-embedding-3-small"),
  embeddingDimension: 1536, // Needs to match your embedding model
});
```

## Add context to RAG

Add content with text chunks. Each call to `add` will create a new **entry**. It
will embed the chunks automatically if you don't provide them.

```ts
export const add = action({
  args: { text: v.string() },
  handler: async (ctx, { text }) => {
    // Add the text to a namespace shared by all users.
    await rag.add(ctx, {
      namespace: "all-users",
      text,
    });
  },
});
```

See below for how to chunk the text yourself or add content asynchronously, e.g.
to handle large files.

## Semantic Search

Search across content with vector similarity

- `text` is a string with the full content of the results, for convenience. It
  is in order of the entries, with titles at each entry boundary, and separators
  between non-sequential chunks. See below for more details.
- `results` is an array of matching chunks with scores and more metadata.
- `entries` is an array of the entries that matched the query. Each result has a
  `entryId` referencing one of these source entries.
- `usage` contains embedding token usage information. Will be `{ tokens: 0 }` if
  no embedding was performed (e.g. when passing pre-computed embeddings).

```ts
export const search = action({
  args: {
    query: v.string(),
  },
  handler: async (ctx, args) => {
    const { results, text, entries, usage } = await rag.search(ctx, {
      namespace: "global",
      query: args.query,
      limit: 10,
      vectorScoreThreshold: 0.5, // Only return results with a score >= 0.5
    });

    return { results, text, entries, usage };
  },
});
```

## Generate a response based on RAG context

Once you have searched for the context, you can use it with an LLM.

Generally you'll already be using something to make LLM requests, e.g. the
[Agent Component](https://www.convex.dev/components/agent), which tracks the
message history for you. See the
[Agent Component docs](https://docs.convex.dev/agents) for more details on doing
RAG with the Agent Component.

However, if you just want a one-off response, you can use the `generateText`
function as a convenience.

This will automatically search for relevant entries and use them as context for
the LLM, using default formatting.

The arguments to `generateText` are compatible with all arguments to
`generateText` from the AI SDK.

```ts
export const askQuestion = action({
  args: {
    prompt: v.string(),
  },
  handler: async (ctx, args) => {
    const userId = await getAuthUserId(ctx);
    const { text, context } = await rag.generateText(ctx, {
      search: { namespace: userId, limit: 10 },
      prompt: args.prompt,
      model: openai.chat("gpt-4o-mini"),
    });
    return { answer: text, context };
  },
```

Note: You can specify any of the search options available on `rag.search`.

## Filtered Search

You can provide filters when adding content and use them to search. To do this,
you'll need to give the RAG component a list of the filter names. You can
optionally provide a type parameter for type safety (no runtime validation).

Note: these filters can be OR'd together when searching. In order to get an AND,
you provide a filter with a more complex value, such as `categoryAndType` below.

```ts
// convex/example.ts
import { components } from "./_generated/api";
import { RAG } from "@convex-dev/rag";
// Any AI SDK model that supports embeddings will work.
import { openai } from "@ai-sdk/openai";

// Optional: Add type safety to your filters.
type FilterTypes = {
  category: string;
  contentType: string;
  categoryAndType: { category: string; contentType: string };
};

const rag = new RAG<FilterTypes>(components.rag, {
  textEmbeddingModel: openai.embedding("text-embedding-3-small"),
  embeddingDimension: 1536, // Needs to match your embedding model
  filterNames: ["category", "contentType", "categoryAndType"],
});
```

Adding content with filters:

```ts
await rag.add(ctx, {
  namespace: "global",
  text,
  filterValues: [
    { name: "category", value: "news" },
    { name: "contentType", value: "article" },
    {
      name: "categoryAndType",
      value: { category: "news", contentType: "article" },
    },
  ],
});
```

Search with metadata filters:

```ts
export const searchForNewsOrSports = action({
  args: {
    query: v.string(),
  },
  handler: async (ctx, args) => {
    const userId = await getUserId(ctx);
    if (!userId) throw new Error("Unauthorized");

    const results = await rag.search(ctx, {
      namespace: userId,
      query: args.query,
      filters: [
        { name: "category", value: "news" },
        { name: "category", value: "sports" },
      ],
      limit: 10,
    });

    return results;
  },
});
```

### Add surrounding chunks to results for context

Instead of getting just the single matching chunk, you can request surrounding
chunks so there's more context to the result.

Note: If there are results that have overlapping ranges, it will not return
duplicate chunks, but instead give priority to adding the "before" context to
each chunk. For example if you requested 2 before and 1 after, and your results
were for the same entryId indexes 1, 4, and 7, the results would be:

```ts
[
  // Only one before chunk available, and leaves chunk2 for the next result.
  { order: 1, content: [chunk0, chunk1], startOrder: 0, ... },
  // 2 before chunks available, but leaves chunk5 for the next result.
  { order: 4, content: [chunk2, chunk3, chunk4], startOrder: 2, ... },
  // 2 before chunks available, and includes one after chunk.
  { order: 7, content: [chunk5, chunk6, chunk7, chunk8], startOrder: 5, ... },
]
```

```ts
export const searchWithContext = action({
  args: {
    query: v.string(),
    userId: v.string(),
  },
  handler: async (ctx, args) => {
    const { results, text, entries, usage } = await rag.search(ctx, {
      namespace: args.userId,
      query: args.query,
      chunkContext: { before: 2, after: 1 }, // Include 2 chunks before, 1 after
      limit: 5,
    });

    return { results, text, entries, usage };
  },
});
```

<!-- END: Include on https://convex.dev/components -->

## Formatting results

Formatting the results for use in a prompt depends a bit on the use case. By
default, the results will be sorted by score, not necessarily in the order they
appear in the original text. You may want to sort them by the order they appear
in the original text so they follow the flow of the original document.

For convenience, the `text` field of the search results is a string formatted
with `...` separating non-sequential chunks, `---` separating entries, and
`# Title:` at each entry boundary (if titles are available).

```ts
const { text } = await rag.search(ctx, { ... });
console.log(text);
```

```txt
## Title 1:
Chunk 1 contents
Chunk 2 contents

...

Chunk 8 contents
Chunk 9 contents

---

## Title 2:
Chunk 4 contents
Chunk 5 contents
```

There is also a `text` field on each entry that is the full text of the entry,
similarly formatted with `...` separating non-sequential chunks, if you want to
format each entry differently.

For a fully custom format, you can use the `results` field and entries directly:

```ts
const { results, text, entries } = await rag.search(ctx, {
  namespace: args.userId,
  query: args.query,
  chunkContext: { before: 2, after: 1 }, // Include 2 chunks before, 1 after
  limit: 5,
  vectorScoreThreshold: 0.5, // Only return results with a score >= 0.5
});

// Get results in the order of the entries (highest score first)
const contexts = entries
  .map((e) => {
    const ranges = results
      .filter((r) => r.entryId === e.entryId)
      .sort((a, b) => a.startOrder - b.startOrder);
    let text = (e.title ?? "") + ":\n\n";
    let previousEnd = 0;
    for (const range of ranges) {
      if (range.startOrder !== previousEnd) {
        text += "\n...\n";
      }
      text += range.content.map((c) => c.text).join("\n");
      previousEnd = range.startOrder + range.content.length;
    }
    return {
      ...e,
      entryId: e.entryId as EntryId,
      filterValues: e.filterValues as EntryFilterValues<FitlerSchemas>[],
      text,
    };
  })
  .map((e) => (e.title ? `# ${e.title}:\n${e.text}` : e.text));

await generateText({
  model: openai.chat("gpt-4o-mini"),
  prompt:
    "Use the following context:\n\n" +
    contexts.join("\n---\n") +
    "\n\n---\n\n Based on the context, answer the question:\n\n" +
    args.query,
});
```

## Using keys to gracefully replace content

When you add content to a namespace, you can provide a `key` to uniquely
identify the content. If you add content with the same key, it will make a new
entry to replace the old one.

```ts
await rag.add(ctx, { namespace: userId, key: "my-file.txt", text });
```

When a new document is added, it will start with a status of "pending" while it
chunks, embeds, and inserts the data into the database. Once all data is
inserted, it will iterate over the chunks and swap the old content embeddings
with the new ones, and then update the status to "ready", marking the previous
version as "replaced".

The old content is kept around by default, so in-flight searches will get
results for old vector search results. See below for more details on deleting.

This means that if searches are happening while the document is being added,
they will see the old content results This is useful if you want to add content
to a namespace and then immediately search for it, or if you want to add content
to a namespace and then immediately add more content to the same namespace.

## Using your own content splitter

By default, the component uses the `defaultChunker` to split the content into
chunks. You can pass in your own content chunks to the `add` or `addAsync`
functions.

```ts
const chunks = await textSplitter.split(content);
await rag.add(ctx, { namespace: "global", chunks });
```

Note: The `textSplitter` here could be LangChain, Mastra, or something custom.
The simplest version makes an array of strings like `content.split("\n")`.

Note: you can pass in an async iterator instead of an array to handle large
content. Or use the `addAsync` function (see below).

## Providing custom embeddings per-chunk

In addition to the text, you can provide your own embeddings for each chunk.

This can be beneficial if you want to embed something other than the chunk
contents, e.g. a summary of each chunk.

```ts
const chunks = await textSplitter.split(content);
const chunksWithEmbeddings = await Promise.all(
  chunks.map(async (chunk) => {
    return {
      ...chunk,
      embedding: await embedSummary(chunk),
    };
  }),
);
await rag.add(ctx, { namespace: "global", chunks });
```

## Add Entries Asynchronously using File Storage

For large files, you can upload them to file storage, then provide a chunker
action to split them into chunks.

In `convex/http.ts`:

```ts
import { corsRouter } from "convex-helpers/server/cors";
import { httpRouter } from "convex/server";
import { internal } from "./_generated/api.js";
import { DataModel } from "./_generated/dataModel.js";
import { httpAction } from "./_generated/server.js";
import { rag } from "./example.js";

const cors = corsRouter(httpRouter());

cors.route({
  path: "/upload",
  method: "POST",
  handler: httpAction(async (ctx, request) => {
    const storageId = await ctx.storage.store(await request.blob());
    await rag.addAsync(ctx, {
      namespace: "all-files",
      chunkerAction: internal.http.chunkerAction,
      onComplete: internal.foo.docComplete, // See next section
      metadata: { storageId },
    });
    return new Response();
  }),
});

export const chunkerAction = rag.defineChunkerAction(async (ctx, args) => {
  const storageId = args.entry.metadata!.storageId;
  const file = await ctx.storage.get(storageId);
  const text = await new TextDecoder().decode(await file!.arrayBuffer());
  return { chunks: text.split("\n\n") };
});

export default cors.http;
```

You can upload files directly to a Convex action, httpAction, or upload url. See
the [docs](https://docs.convex.dev/file-storage/upload-files) for details.

### OnComplete Handling

You can register an `onComplete` handler when adding content that will be called
when the entry was created and is ready, or if there was an error or it was
replaced before it finished.

```ts
// in an action
await rag.add(ctx, { namespace, text, onComplete: internal.foo.docComplete });

// in convex/foo.ts
export const docComplete = rag.defineOnComplete<DataModel>(
  async (ctx, { replacedEntry, entry, namespace, error }) => {
    if (error) {
      await rag.delete(ctx, { entryId: entry.entryId });
      return;
    }
    if (replacedEntry) {
      await rag.delete(ctx, { entryId: replacedEntry.entryId });
    }
    // You can associate the entry with your own data here. This will commit
    // in the same transaction as the entry becoming ready.
  },
);
```

Note: The `onComplete` callback is only triggered when new content is processed.
If you add content that already exists (`contentHash` did not change for the
same `key`), `onComplete` will not be called. To handle this case, you can check
the return value of `rag.add()`:

```ts
const { status, created } = await rag.add(ctx, {
  namespace,
  text,
  key: "my-key",
  contentHash: "...",
  onComplete: internal.foo.docComplete,
});

if (status === "ready" && !created) {
  // Entry already existed - onComplete will not be called
  // Handle this case if needed
}
```

### Add Entries with filters from a URL

Here's a simple example fetching content from a URL to add.

It also adds filters to the entry, so you can search for it later by category,
contentType, or both.

```ts
export const add = action({
  args: { url: v.string(), category: v.string() },
  handler: async (ctx, { url, category }) => {
    const response = await fetch(url);
    const content = await response.text();
    const contentType = response.headers.get("content-type");

    const { entryId } = await rag.add(ctx, {
      namespace: "global", // namespace can be any string
      key: url,
      chunks: content.split("\n\n"),
      filterValues: [
        { name: "category", value: category },
        { name: "contentType", value: contentType },
        // To get an AND filter, use a filter with a more complex value.
        { name: "categoryAndType", value: { category, contentType } },
      ],
    });

    return { entryId };
  },
});
```

### Lifecycle Management

You can delete the old content by calling `rag.delete` with the entryId of the
old version.

Generally you'd do this:

1. When using `rag.add` with a key returns a `replacedEntry`.
1. When your `onComplete` handler provides a non-null `replacedEntry` argument.
1. Periodically by querying:

```ts
// in convex/crons.ts
import { cronJobs } from "convex/server";
import { internal } from "./_generated/api.js";
import { internalMutation } from "./_generated/server.js";
import { v } from "convex/values";
import { rag } from "./example.js";
import { assert } from "convex-helpers";

const WEEK = 7 * 24 * 60 * 60 * 1000;

export const deleteOldContent = internalMutation({
  args: { cursor: v.optional(v.string()) },
  handler: async (ctx, args) => {
    const toDelete = await rag.list(ctx, {
      status: "replaced",
      paginationOpts: { cursor: args.cursor ?? null, numItems: 100 },
    });

    for (const entry of toDelete.page) {
      assert(entry.status === "replaced");
      if (entry.replacedAt >= Date.now() - WEEK) {
        return; // we're done when we catch up to a week ago
      }
      await rag.delete(ctx, { entryId: entry.entryId });
    }
    if (!toDelete.isDone) {
      await ctx.scheduler.runAfter(0, internal.example.deleteOldContent, {
        cursor: toDelete.continueCursor,
      });
    }
  },
});

// See example/convex/crons.ts for a complete example.
const crons = cronJobs();
crons.interval(
  "deleteOldContent",
  { hours: 1 },
  internal.crons.deleteOldContent,
  {},
);
export default crons;
```

## Working with types

You can use the provided types to validate and store data.
`import { ... } from "@convex-dev/rag";`

Types for the various elements:

`Entry`, `EntryFilter`, `SearchEntry`, `SearchResult`

- `SearchEntry` is an `Entry` with a `text` field including the combined search
  results for that entry, whereas a `SearchResult` is a specific chunk result,
  along with surrounding chunks.

`EntryId`, `NamespaceId`

- While the `EntryId` and `NamespaceId` are strings under the hood, they are
  given more specific types to make it easier to use them correctly.

Validators can be used in `args` and schema table definitions: `vEntry`,
`vEntryId`, `vNamespaceId`, `vSearchEntry`, `vSearchResult`

e.g. `defineTable({ myDocTitle: v.string(), entryId: vEntryId })`

The validators for the branded IDs will only validate they are strings, but will
have the more specific types, to provide type safety.

## Utility Functions

In addition to the function on the `rag` instance, there are other utilities
provided:

### `defaultChunker`

This is the default chunker used by the `add` and `addAsync` functions.

It is customizable, but by default:

- It tries to break up the text into paragraphs between 100-1k characters.
- It will combine paragraphs to meet the minimum character count (100).
- It will break up paragraphs into separate lines to keep it under 1k.
- It will not split up a single line unless it's longer than 10k characters.

```ts
import { defaultChunker } from "@convex-dev/rag";

const chunks = defaultChunker(text, {
  // these are the defaults
  minLines: 1,
  minCharsSoftLimit: 100,
  maxCharsSoftLimit: 1000,
  maxCharsHardLimit: 10000,
  delimiter: "\n\n",
});
```

### `hybridRank`

This is an implementation of "Reciprocal Rank Fusion" for ranking search results
based on multiple scoring arrays. The premise is that if both arrays of results
are sorted by score, the best results show up near the top of both arrays and
should be preferred over results higher in one but much lower in the other.

```ts
import { hybridRank } from "@convex-dev/rag";

const textSearchResults = [id1, id2, id3];
const vectorSearchResults = [id2, id3, id1];
const results = hybridRank([textSearchResults, vectorSearchResults]);
// results = [id2, id1, id3]
```

It can take more than two arrays, and you can provide weights for each array.

```ts
const recentSearchResults = [id5, id4, id3];
const results = hybridRank(
  [textSearchResults, vectorSearchResults, recentSearchResults],
  {
    weights: [2, 1, 3], // prefer recent results more than text or vector
  },
);
// results = [ id3, id5, id1, id2, id4 ]
```

To have it more biased towards the top few results, you can set the `k` value to
a lower number (10 by default).

```ts
const results = hybridRank(
  [textSearchResults, vectorSearchResults, recentSearchResults],
  { k: 1 },
);
// results = [ id5, id1, id3, id2, id4 ]
```

### `contentHashFromArrayBuffer`

This generates the hash of a file's contents, which can be used to avoid adding
the same file twice.

Note: doing `blob.arrayBuffer()` will consume the blob's data, so you'll need to
make a new blob to use it after calling this function.

```ts
import { contentHashFromArrayBuffer } from "@convex-dev/rag";

export const addFile = action({
  args: { bytes: v.bytes() },
  handler: async (ctx, { bytes }) => {
    const hash = await contentHashFromArrayBuffer(bytes);

    const existing = await rag.findEntryByContentHash(ctx, {
      namespace: "global",
      key: "my-file.txt",
      contentHash: hash,
    });
    if (existing) {
      console.log("File contents are the same, skipping");
      return;
    }
    const blob = new Blob([bytes], { type: "text/plain" });
    //...
  },
});
```

### `guessMimeTypeFromExtension`

This guesses the mime type of a file from its extension.

```ts
import { guessMimeTypeFromExtension } from "@convex-dev/rag";

const mimeType = guessMimeTypeFromExtension("my-file.mjs");
console.log(mimeType); // "text/javascript"
```

### `guessMimeTypeFromContents`

This guesses the mime type of a file from the first few bytes of its contents.

```ts
import { guessMimeTypeFromContents } from "@convex-dev/rag";

const mimeType = guessMimeTypeFromContents(await file.arrayBuffer());
```

### Example Usage

See more example usage in [example.ts](./example/convex/example.ts).

### Running the example

Run the example with `npm run setup && npm run dev`.