Skip to content

Feature/qdrant mrl support #6057

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

Closed
voreskovic wants to merge 2 commits into from
+217 −20
Closed

Feature/qdrant mrl support #6057

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
2e04843
test
Mar 25, 2026
aa1e43d
Added support for Matryoshka Representation Learning (MRL) Qdrant bas…
Mar 25, 2026
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
2 changes: 2 additions & 0 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,8 @@

<div align="center">

Nice
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Mar 25, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

This line appears to be an accidental addition and should be removed to keep the README clean.


[![Release Notes](https://img.shields.io/github/release/FlowiseAI/Flowise)](https://github.com/FlowiseAI/Flowise/releases)
[![Discord](https://img.shields.io/discord/1087698854775881778?label=Discord&logo=discord)](https://discord.gg/jbaHfsRVBW)
[![Twitter Follow](https://img.shields.io/twitter/follow/FlowiseAI?style=social)](https://twitter.com/FlowiseAI)
Expand Down
235 changes: 215 additions & 20 deletions packages/components/nodes/vectorstores/Qdrant/Qdrant.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,71 @@ type QdrantAddDocumentOptions = {
ids?: string[]
}

/**
* Parse the MRL dimensions string into an array of integers.
* E.g. "512, 256, 128" -> [512, 256, 128]
*/
function parseMrlDimensions(raw: string): number[] {
return raw
.split(',')
.map((s) => parseInt(s.trim(), 10))
.filter((n) => !isNaN(n) && n > 0)
}

/**
* Build Qdrant named-vectors config for MRL.
* Returns e.g. { "full": { size: 1024, distance: "Cosine" }, "512": { size: 512, distance: "Cosine" } }
* Truncated vectors use their dimension as the vector name (e.g. "512", "256").
*/
function buildMrlVectorsConfig(fullDimension: number, truncatedDimensions: number[], distance: string): Record<string, any> {
const vectors: Record<string, any> = {
full: { size: fullDimension, distance }
}
for (const dim of truncatedDimensions) {
if (dim >= fullDimension) continue
vectors[`${dim}`] = { size: dim, distance }
}
return vectors
}

/**
* Truncate an embedding vector to the specified length and L2-normalize it.
* MRL embeddings retain semantic meaning when truncated, but re-normalization
* is recommended for cosine similarity to remain well-behaved.
*/
function truncateAndNormalize(embedding: number[], targetDim: number): number[] {
const truncated = embedding.slice(0, targetDim)
const norm = Math.sqrt(truncated.reduce((sum, v) => sum + v * v, 0))
if (norm === 0) return truncated
return truncated.map((v) => v / norm)
}

/**
* Build the named-vector map for a single point during MRL upsert.
*/
function buildMrlVectorMap(embedding: number[], truncatedDimensions: number[], fullDimension: number): Record<string, number[]> {
const vectorMap: Record<string, number[]> = {
full: embedding
}
for (const dim of truncatedDimensions) {
if (dim >= fullDimension) continue
vectorMap[`${dim}`] = truncateAndNormalize(embedding, dim)
}
return vectorMap
}

/**
* Ensure the MRL collection exists with named vectors.
* If the collection already exists, this is a no-op.
*/
async function ensureMrlCollection(client: QdrantClient, collectionName: string, collectionConfig: Record<string, any>): Promise<void> {
const response = await client.getCollections()
const collectionNames = response.collections.map((c: any) => c.name)
if (!collectionNames.includes(collectionName)) {
await client.createCollection(collectionName, collectionConfig)
}
}
Comment on lines +76 to +82
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Mar 25, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The current implementation of ensureMrlCollection fetches all collections to check if one exists, which can be inefficient if there are many collections. A more performant approach is to try fetching the specific collection by name and creating it only if a "not found" error (e.g., status 404) is returned.

async function ensureMrlCollection(client: QdrantClient, collectionName: string, collectionConfig: Record<string, any>): Promise<void> {
    try {
        await client.getCollection(collectionName)
    } catch (e: any) {
        if (e.status === 404) {
            await client.createCollection(collectionName, collectionConfig)
        } else {
            throw e
        }
    }
}


class Qdrant_VectorStores implements INode {
label: string
name: string
Expand All @@ -33,7 +98,7 @@ class Qdrant_VectorStores implements INode {
constructor() {
this.label = 'Qdrant'
this.name = 'qdrant'
this.version = 5.0
this.version = 6.0
this.type = 'Qdrant'
this.icon = 'qdrant.png'
this.category = 'Vector Stores'
Expand Down Expand Up @@ -148,6 +213,37 @@ class Qdrant_VectorStores implements INode {
],
additionalParams: true
},
{
label: 'Enable MRL (Matryoshka Representation Learning)',
name: 'mrlEnabled',
description:
'Store embeddings at multiple truncated dimensions using Qdrant named vectors. Requires an MRL-trained embedding model (e.g. OpenAI text-embedding-3-*, Nomic Embed, jina-embeddings-v2).',
type: 'boolean',
default: false,
additionalParams: true,
optional: true
},
{
label: 'MRL Truncated Dimensions',
name: 'mrlDimensions',
description:
'Comma-separated list of truncated dimensions to store alongside the full embedding. E.g. "512, 256, 128". Each must be smaller than the full Vector Dimension.',
type: 'string',
placeholder: '512, 256',
additionalParams: true,
optional: true
},
{
label: 'MRL Search Vector',
name: 'mrlSearchVector',
description:
'Which named vector to use for similarity search. "full" uses the original embedding, or specify a dimension number (e.g. "256") to search the truncated vector. Smaller dimensions are faster but less precise.',
type: 'string',
default: 'full',
placeholder: 'full',
additionalParams: true,
optional: true
},
{
label: 'Additional Collection Cofiguration',
name: 'qdrantCollectionConfiguration',
Expand Down Expand Up @@ -205,6 +301,10 @@ class Qdrant_VectorStores implements INode {
const metadataPayloadKey = nodeData.inputs?.metadataPayloadKey || 'metadata'
const isFileUploadEnabled = nodeData.inputs?.fileUpload as boolean

const mrlEnabled = nodeData.inputs?.mrlEnabled as boolean
const mrlDimensionsRaw = (nodeData.inputs?.mrlDimensions as string) || ''
const mrlDimensions = mrlEnabled ? parseMrlDimensions(mrlDimensionsRaw) : []

const credentialData = await getCredentialData(nodeData.credential ?? '', options)
const qdrantApiKey = getCredentialParam('qdrantApiKey', credentialData, nodeData)

Expand All @@ -217,7 +317,7 @@ class Qdrant_VectorStores implements INode {
})

const flattenDocs = docs && docs.length ? flatten(docs) : []
const finalDocs = []
const finalDocs: Document[] = []
for (let i = 0; i < flattenDocs.length; i += 1) {
if (flattenDocs[i] && flattenDocs[i].pageContent) {
if (isFileUploadEnabled && options.chatId) {
Expand All @@ -227,24 +327,32 @@ class Qdrant_VectorStores implements INode {
}
}

const fullDimension = qdrantVectorDimension ? parseInt(qdrantVectorDimension, 10) : 1536
const distance = qdrantSimilarity ?? 'Cosine'

const collectionConfig =
mrlEnabled && mrlDimensions.length > 0
? { vectors: buildMrlVectorsConfig(fullDimension, mrlDimensions, distance) }
: { vectors: { size: fullDimension, distance } }
Comment on lines +333 to +336
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Mar 25, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The condition mrlEnabled && mrlDimensions.length > 0 is used multiple times. To improve readability and reduce repetition, consider storing this boolean check in a variable like isMrlActive and reusing it.

References
  1. Prioritize code readability and understandability over conciseness. A series of simple, chained operations can be preferable to a single, more complex one if it improves understandability and reduces the potential for future errors.


const dbConfig: QdrantLibArgs = {
client: client as any,
url: qdrantServerUrl,
collectionName,
collectionConfig: {
vectors: {
size: qdrantVectorDimension ? parseInt(qdrantVectorDimension, 10) : 1536,
distance: qdrantSimilarity ?? 'Cosine'
}
},
collectionConfig,
contentPayloadKey,
metadataPayloadKey
}

try {
if (recordManager) {
const vectorStore = new QdrantVectorStore(embeddings, dbConfig)
await vectorStore.ensureCollection()

if (mrlEnabled && mrlDimensions.length > 0) {
await ensureMrlCollection(client, collectionName, collectionConfig)
} else {
await vectorStore.ensureCollection()
}
Comment on lines +351 to +355
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Mar 25, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

This call to ensure the collection exists is redundant. The overridden addVectors method on line 357 also calls ensureMrlCollection (or vectorStore.ensureCollection). Since the index function will call addVectors, the check for the collection's existence should only be within addVectors. Please remove this redundant block.


vectorStore.addVectors = async (
vectors: number[][],
Expand All @@ -255,11 +363,18 @@ class Qdrant_VectorStores implements INode {
return
}

await vectorStore.ensureCollection()
if (mrlEnabled && mrlDimensions.length > 0) {
await ensureMrlCollection(client, collectionName, collectionConfig)
} else {
await vectorStore.ensureCollection()
}

const points = vectors.map((embedding, idx) => ({
id: documentOptions?.ids?.length ? documentOptions?.ids[idx] : uuid(),
vector: embedding,
vector:
mrlEnabled && mrlDimensions.length > 0
? buildMrlVectorMap(embedding, mrlDimensions, fullDimension)
: embedding,
payload: {
[contentPayloadKey]: documents[idx].pageContent,
[metadataPayloadKey]: documents[idx].metadata,
Expand All @@ -274,13 +389,13 @@ class Qdrant_VectorStores implements INode {
const batchPoints = points.slice(i, i + batchSize)
await client.upsert(collectionName, {
wait: true,
points: batchPoints
points: batchPoints as any
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Mar 25, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

Casting points to any compromises type safety. You can import PointStruct from @qdrant/js-client-rest and apply it to your points array. This will provide better type checking and allow you to remove the as any cast. This feedback also applies to lines 398, 457, and 463.

})
}
} else {
await client.upsert(collectionName, {
wait: true,
points
points: points as any
})
}
} catch (e: any) {
Expand Down Expand Up @@ -318,16 +433,50 @@ class Qdrant_VectorStores implements INode {

return res
} else {
if (_batchSize) {
const batchSize = parseInt(_batchSize, 10)
for (let i = 0; i < finalDocs.length; i += batchSize) {
const batch = finalDocs.slice(i, i + batchSize)
await QdrantVectorStore.fromDocuments(batch, embeddings, dbConfig)
if (mrlEnabled && mrlDimensions.length > 0) {
await ensureMrlCollection(client, collectionName, collectionConfig)

const texts = finalDocs.map((doc) => doc.pageContent)
const allEmbeddings = await embeddings.embedDocuments(texts)

const points = allEmbeddings.map((embedding: number[], idx: number) => ({
id: uuid(),
vector: buildMrlVectorMap(embedding, mrlDimensions, fullDimension),
payload: {
[contentPayloadKey]: finalDocs[idx].pageContent,
[metadataPayloadKey]: finalDocs[idx].metadata
}
}))

if (_batchSize) {
const batchSize = parseInt(_batchSize, 10)
for (let i = 0; i < points.length; i += batchSize) {
const batchPoints = points.slice(i, i + batchSize)
await client.upsert(collectionName, {
wait: true,
points: batchPoints as any
})
}
} else {
await client.upsert(collectionName, {
wait: true,
points: points as any
})
}

return { numAdded: finalDocs.length, addedDocs: finalDocs }
} else {
await QdrantVectorStore.fromDocuments(finalDocs, embeddings, dbConfig)
if (_batchSize) {
const batchSize = parseInt(_batchSize, 10)
for (let i = 0; i < finalDocs.length; i += batchSize) {
const batch = finalDocs.slice(i, i + batchSize)
await QdrantVectorStore.fromDocuments(batch, embeddings, dbConfig)
}
} else {
await QdrantVectorStore.fromDocuments(finalDocs, embeddings, dbConfig)
}
return { numAdded: finalDocs.length, addedDocs: finalDocs }
}
Comment on lines +436 to 479
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Mar 25, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

There is significant code duplication in this block. The logic for batch upserting points is repeated from the addVectors override (lines 386-400). Additionally, the return { numAdded: finalDocs.length, addedDocs: finalDocs } statement is duplicated in both branches of the MRL conditional.

To improve maintainability, consider extracting the batching logic into a helper function and moving the common return statement outside the conditional block.

References
  1. Prioritize code readability and understandability over conciseness. A series of simple, chained operations can be preferable to a single, more complex one if it improves understandability and reduces the potential for future errors.

return { numAdded: finalDocs.length, addedDocs: finalDocs }
}
} catch (e) {
throw new Error(e)
Expand Down Expand Up @@ -416,6 +565,9 @@ class Qdrant_VectorStores implements INode {
const metadataPayloadKey = nodeData.inputs?.metadataPayloadKey || 'metadata'
const isFileUploadEnabled = nodeData.inputs?.fileUpload as boolean

const mrlEnabled = nodeData.inputs?.mrlEnabled as boolean
const mrlSearchVector = (nodeData.inputs?.mrlSearchVector as string) || 'full'

const k = topK ? parseFloat(topK) : 4

const credentialData = await getCredentialData(nodeData.credential ?? '', options)
Expand Down Expand Up @@ -480,6 +632,49 @@ class Qdrant_VectorStores implements INode {

const vectorStore = await QdrantVectorStore.fromExistingCollection(embeddings, dbConfig)

// For MRL: override similaritySearchVectorWithScore to use the `using` parameter
if (mrlEnabled && mrlSearchVector) {
const searchVectorName = mrlSearchVector
const origFilter = retrieverConfig.filter

vectorStore.similaritySearchVectorWithScore = async (
query: number[],
searchK: number,
filter?: Record<string, any>
): Promise<[Document, number][]> => {
const searchFilter = filter ?? origFilter

// Truncate query vector if searching a smaller named vector
let searchQuery: number[] = query
if (searchVectorName !== 'full') {
const targetDim = parseInt(searchVectorName, 10)
if (!isNaN(targetDim) && query.length > targetDim) {
searchQuery = truncateAndNormalize(query, targetDim)
}
}

const results = (
await client.query(collectionName, {
query: searchQuery,
using: searchVectorName,
limit: searchK,
filter: searchFilter,
with_payload: [metadataPayloadKey, contentPayloadKey],
with_vector: false
})
).points

return results.map((res: any) => [
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Mar 25, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

Using any for the res parameter reduces type safety. You can import ScoredPoint from @qdrant/js-client-rest and use it to type the res parameter. This will improve code quality and provide better type-checking.

Suggested change
return results.map((res: any) => [
return results.map((res: ScoredPoint) => [

new Document({
id: res.id,
metadata: res.payload?.[metadataPayloadKey],
pageContent: res.payload?.[contentPayloadKey] ?? ''
}),
res.score
])
}
}

if (output === 'retriever') {
const retriever = vectorStore.asRetriever(retrieverConfig)
return retriever
Expand Down