How to confirm that a knowledge graph has been used in the retrieval

[hcxx-x]

How can I test that my retrieval uses a knowledge graph? I have already generated a knowledge graph, but the content of the retrieval test remains the same whether the knowledge graph is turned on or off

[dosubot[bot]]

Hi @hcxx-x! I'm Dosu and I'm helping the InfiniFlow team.

When knowledge graph retrieval is working, the results should be visibly different — KG results are inserted at the top of the result list with similarity: 1.0 and are labeled with "Related content in Knowledge Graph" as the document name [1]. They also contain structured entity/relation tables rather than plain document chunks [2].

If you're seeing no difference with KG on/off, the most common cause is a configuration timing issue: knowledge graph settings are locked at file import time. If you enabled knowledge graph settings after uploading your files, you need to delete those files and re-import them with KG already enabled [3].

Here's the full required workflow (all 4 steps must be completed in order):

1. Configure entity types in the dataset's Configuration page before uploading files
2. Parse — upload and parse files (they must produce visible chunks successfully)
3. Generate — click the "Generate" button on the Files page and select "Knowledge graph"
4. Enable in retrieval — toggle "Use Knowledge Graph in search" when querying

If any step is missed or done out of order, KG won't be utilized during retrieval [3].

Also note: if you're running v0.25.0, there was a bug where KG generation failed silently when files had zero chunks. This was fixed in v0.25.1 via PRs #14383 and #14459 [3]. Upgrading may help if you're on the older version.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.

[rishavsunny12]

you can use tools like these to evaluate your graph
https://github.com/rishavsunny12/ragProbe

[Ganesh-403]

@hcxx-x A simple way to verify that the Knowledge Graph is actually being used is through Retrieval Testing.

When Use knowledge graph is enabled, the retrieval results should be different from standard hybrid retrieval:

• The retrieved items will include entity descriptions, relationship descriptions, and the top community report from the knowledge graph, rather than only document chunks.
• Knowledge graph retrieval uses a different retrieval pipeline (entity extraction → graph traversal → community retrieval), so you should also notice that retrieval takes a bit longer because it invokes the configured LLM to analyze the query. :contentReference[oaicite:0]{index=0}

A few things to check:

1. Use a query that requires multi-hop reasoning.

   For example, instead of asking:

   "What is Product A?"

   try something like:

   "Which suppliers provide Product A, and which countries are they associated with?"

   Knowledge graphs are most beneficial for questions involving relationships between entities. A simple factual lookup may produce nearly identical results with or without KG enabled.

2. Verify that the graph was actually constructed.

   Under your dataset, you should see a Knowledge graph entry after generation completes. If the graph wasn't successfully generated, enabling the toggle won't have any effect. :contentReference[oaicite:1]{index=1}

3. Compare Retrieval Test with KG on vs. off.

   Run the exact same query twice:

   • once with Use knowledge graph = Off
   • once with Use knowledge graph = On

   If the retrieved content and latency are identical, then either:

   • the query doesn't benefit from graph traversal,
   • the graph contains little additional relationship information,
   • or the KG isn't being consulted as expected.

Could you also share:

• Which RAGFlow version you're using?
• What kind of documents are in the dataset (technical docs, manuals, contracts, etc.)?
• An example query where you expected the knowledge graph to make a difference?

That would help determine whether this is expected behavior or potentially a bug.

If this solves your problem, feel free to mark it as the accepted answer so others can find it easily.