Dynamic Knowledge Retrieval

Author: xiaozie-xy
Version: 0.0.1
Type: Tool plugin

Overview

Dynamic Knowledge Retrieval is a Dify tool plugin. It searches the knowledge base identified by a dataset ID passed in at runtime.

The built-in knowledge retrieval node stores dataset IDs in the workflow graph, so one run cannot switch to another dataset. This plugin adds a parameter. Pass the knowledge base ID that matches the current task.

Features

Runtime dataset selection. Each run searches the knowledge base given by .
Search method. Follow the dataset setting, or override it with semantic, full-text, keyword, or hybrid search.
Result size and score. Set top K and an optional score threshold.
Rerank. Pick a rerank model configured in the workspace. When rerank is off, the dataset setting is kept.
Hybrid weights. Follow the saved dataset weights, or use semantic-first, keyword-first, or custom weights. Embedding model names come from the dataset.
Metadata filters. Optional. Pass conditions as a JSON array. Each item has a field name, comparison operator, and value. Combine them with and or or. Leave the array empty to skip filtering.
Attachments. Optionally pass attachment IDs for multimodal retrieval.
Two outputs. is the chunk list. is the joined text.

Installation

Upload on the Dify plugins page. If a self-hosted instance rejects the unsigned local package, turn off forced plugin signature checks in the deployment configuration and install it again. Then enter the knowledge API base URL and an API key scoped to all knowledge bases.

Configuration

Provider credentials

Configure these once. This key is not an application API key.

Parameter	Description	Required
	Knowledge service API base URL, including . Cloud: . From the plugin daemon in Docker Compose, usually .	Yes
	Create it under Knowledge → Service API. Scope it to all knowledge bases when each run may search a different dataset. A key scoped to one dataset can access only that dataset.	Yes

Tool parameters

Bound from workflow variables

Parameter	Description	Required
	UUID of the knowledge base to search. Pass the ID that matches the current task.	Yes
	Search text, up to 250 characters. Put only what to look up here. Leave formatting or extraction instructions to the downstream LLM node.	Yes
	Optional comma-separated attachment IDs. Example: . Leave empty for text-only search.	No
	Optional JSON string. Example: . Operators are listed below. Leave empty to skip metadata filtering.	No

Set on the tool node

Parameter	Description	Default
	Follow the dataset, or use semantic, full-text, keyword, or hybrid search.	Follow dataset
	Maximum number of chunks, from 1 to 50.	4
	Off keeps the dataset score setting. On uses the threshold below.	Off
	Score threshold from 0 to 1. Used only when the threshold is enabled.	None
	Off keeps the dataset rerank setting. On uses the selected model.	Off
	A rerank model configured in this workspace.	None
	Follow dataset, semantic first, keyword first, or custom.	Follow dataset
	Semantic weight from 0 to 1 when weights are custom. Keyword weight is 1 minus this value.	0.7
	How to combine conditions: and or or.	and

Screenshots

Bind the dataset ID and query for the current run. Leave attachment IDs and metadata conditions empty when they are not needed. Search method follows the dataset by default.

Top K, score threshold, the rerank model, and hybrid weights are set in the node's retrieval settings.

Examples

Search one knowledge base for the current task:

Include two attachments:

Hybrid search with custom weights and two metadata conditions:

Semantic first uses semantic weight 0.7 and keyword weight 0.3. Keyword first reverses those values. When rerank is off, the plugin sends weighted-score mode.

Metadata comparison operators

Field definitions follow the official docs: metadata filtering conditions.

must be one of the values below.

Operator	Meaning	Example
	Contains the text
	Does not contain the text
	Starts with the text
	Ends with the text
	Equals
	Does not equal
	Is empty. may be omitted
	Is not empty. may be omitted
	Is one of the values. Separate items with commas
	Is none of the values
	Numeric comparison. must be a number
	Earlier than
	Later than

Workflow

Creating, uploading, waiting for indexing, and deleting datasets stay with the caller. The target dataset's indexing status should be completed before this tool runs.

Outputs

Output	Description
	Array of chunks. Each item has , , , , , , , , , , , , , , and . is the document's own metadata. is the parent chunk. Child chunks stay in . and are null when the segment has none.
	Chunk values joined with blank lines, ready for an LLM prompt.
Text message	Same as . When nothing matches, the text says that no chunks were retrieved.

Limitations

The query is limited to 250 characters.
Retrieval can return nothing, or the API can report that the dataset is still initializing, while indexing is incomplete.
Hybrid search needs saved dataset weights, rerank enabled in the plugin, or a hybrid weight choice.
Custom weights replace the dataset setting only when the search method is hybrid.
This plugin does not create, upload to, or delete knowledge bases.
The retrieve API records the query on the knowledge base hit-testing page.

Requirements

Dify 1.17.1. Developed against this version. Test other versions yourself.
Python 3.12

License

Apache License 2.0.

Dynamic Knowledge Retrieval

Dynamic Knowledge Retrieval

Overview

Features

Installation

Configuration

Provider credentials

Tool parameters

Screenshots

Examples

Metadata comparison operators

Workflow

Outputs

Limitations

Requirements

License

Links