Qdrant plugin

The Qdrant plugin provides Genkit indexer and retriever implementations in JS and Go that use the Qdrant.

Installation

npm i genkitx-qdrant

Configuration

To use this plugin, specify it when you call configureGenkit():

import { qdrant } from 'genkitx-qdrant';

const ai = genkit({
  plugins: [
    qdrant([
      {
        embedder: googleAI.embedder('text-embedding-004'),
        collectionName: 'collectionName',
        clientParams: {
          url: 'http://localhost:6333',
        },
      },
    ]),
  ],
});

You'll need to specify a collection name, the embedding model you want to use and the Qdrant client parameters. In addition, there are a few optional parameters:

• embedderOptions: Additional options to pass options to the embedder:

  embedderOptions: { taskType: 'RETRIEVAL_DOCUMENT' },

• contentPayloadKey: Name of the payload filed with the document content. Defaults to "content".

  contentPayloadKey: 'content';

• metadataPayloadKey: Name of the payload filed with the document metadata. Defaults to "metadata".

  metadataPayloadKey: 'metadata';

• dataTypePayloadKey: Name of the payload filed with the document datatype. Defaults to "_content_type".

  dataTypePayloadKey: '_datatype';

• collectionCreateOptions: Additional options when creating the Qdrant collection.

Usage

Import retriever and indexer references like so:

import { qdrantIndexerRef, qdrantRetrieverRef } from 'genkitx-qdrant';

Then, pass the references to retrieve() and index():

// To export an indexer:
export const qdrantIndexer = qdrantIndexerRef('collectionName', 'displayName');

// To export a retriever:
export const qdrantRetriever = qdrantRetrieverRef(
  'collectionName',
  'displayName',
);

You can refer to Retrieval-augmented generation for a general discussion on indexers and retrievers.

Score boosting / fusion

The retriever options accept optional query and prefetch fields, forwarded to the Qdrant Query API. When query is set, the embedded vector becomes a prefetch (overridable via prefetch) and query reranks the candidates, enabling formula boosting, fusion (RRF), and multi-stage prefetch. Omit both for a plain vector query.

const docs = await ai.retrieve({
  retriever: qdrantRetriever,
  query: 'wireless headphones',
  options: {
    k: 20,
    // Boost in-stock items.
    query: { formula: { sum: ['$score', { mult: [0.2, 'in_stock'] }] } },
  },
});

Grouped retrieval (diversity)

Set the optional groupBy field (an indexed payload field) to use the Qdrant Grouping API. Qdrant returns up to groupSize hits per group (default 3) across k groups, so one over-represented group (a large document split into many chunks, say) can't crowd out the rest. The retriever flattens the groups in order and writes each group key to the document's _group metadata, so keep the returned order. Re-sorting by _similarityScore interleaves the groups.

const docs = await ai.retrieve({
  retriever: qdrantRetriever,
  query: 'fire-rated pipe penetration',
  options: {
    k: 8, // up to 8 groups
    groupBy: 'metadata.productFamily',
    groupSize: 3, // up to 3 chunks per family
    filter: {
      must: [{ key: 'metadata.wallType', match: { value: 'flexible' } }],
    },
  },
});