[BUG]LIGHTRAG BUG

[Meteor-de]

Description

When the LightRAG retriever is enabled, the chat session's context appears to get "stuck" on the first document it interacts with. After successfully uploading and indexing two separate documents (Doc A and Doc B), if I start a new chat and select Doc B, the application still loads the knowledge graph from Doc A and generates answers based on Doc A's content.
This indicates that the session state, specifically the file_id used by the LightRAGRetrieverPipeline, is not being correctly updated or reset when a user initiates a new conversation with a different document.
If I ask a question that is highly relevant to the selected document, the system can provide a relatively accurate answer. In this case, a partial knowledge graph is generated, but the main structure of the graph still belongs to the previous document.
Furthermore, for some other documents, the system completely fails to provide any answer or display any graph at all.

在使用 LightRAG 检索器时，聊天会话的上下文似乎被“卡”在了它交互的第一个文档上。我分别上传并成功索引了两个不同的文档（文档A和文档B）后，如果我开启一个新聊天并选择文档B，应用程序仍然会加载文档A的知识图谱，并且基于文档A的内容进行回答。
这表明会话的状态，特别是 LightRAGRetrieverPipeline 使用的 file_id，在用户开启一个新对话并选择不同文档时，没有被正确地更新或重置。
如果选择正确的问题描述，他会给出比较准确的回答，会有部分图的产生，但是图的大体仍然是之前的。并且部分文件仍然完全无法给出回答和显示图。

已尝试的解决方法
1.重启应用程序 (python app.py)。
2.重新启动机器。
3.删除 ktem_app_data 目录以强制进行彻底的数据重置。

Reproduction steps

1.启用 LightRAG 检索器并启动应用
2.上传第一个文档，我们称之为 文档A.pdf。等待索引完成。
3.上传第二个不同的文档，文档B.pdf。等待其索引完成。
4.与 文档A.pdf 开始一个聊天会话，提出一个问题。此时知识图谱显示正确，回答也正确。
5.点击“新聊天”按钮，开启一个全新的对话。
6.在新的聊天窗口中，从文件列表中选择 文档B.pdf。
7.提出了一个与 文档B.pdf 内容相关的问题。

1.Launch the application with the LightRAG retriever enabled.
2.Upload the first document, Document_A.pdf, and wait for indexing to complete.
3.Upload a second, different document, Document_B.pdf, and wait for it to be indexed.
4.Start a chat session with Document_A.pdf and ask a question. The knowledge graph and the 5.answer are both correct at this point.
6.Click the "New Chat" button to start a fresh conversation.
7.In the new chat window, select Document_B.pdf from the file list.
8.Ask a question related to the content of Document_B.pdf.

Screenshots

![DESCRIPTION](LINK.png)

Logs

Retrievers [DocumentRetrievalPipeline(DS=<kotaemon.storages.docstores.lancedb.LanceDBDocumentStore object at 0x7f47604de980>, FSPath=PosixPath('/root/kotaemon/ktem_app_data/user_data/files/index_1'), Index=<class 'ktem.index.file.index.IndexTable'>, Source=<class 'ktem.index.file.index.Source'>, VS=<kotaemon.storages.vectorstores.chroma.ChromaVectorStore object at 0x7f47604dde40>, get_extra_table=False, llm_scorer=LLMTrulensScoring(concurrent=True, normalize=10, prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7f47cf2bc3d0>, system_prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7f47cf2bf520>, top_k=3, user_prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7f47cf2be9b0>), mmr=False, rerankers=[CohereReranking(base_url=None, cohere_api_key='<COHERE_API_KEY>', model_name='rerank-multilingual-v2.0')], retrieval_mode='hybrid', top_k=10, user_id='92021f5bde6949cca8dc7a6b2e983844'), GraphRAGRetrieverPipeline(DS=<theflow.base.unset_ object at 0x7f48325f0e20>, FSPath=<theflow.base.unset_ object at 0x7f48325f0e20>, Index=<class 'ktem.index.file.index.IndexTable'>, Source=<theflow.base.unset_ object at 0x7f48325f0e20>, VS=<theflow.base.unset_ object at 0x7f48325f0e20>, file_ids=[], user_id=<theflow.base.unset_ object at 0x7f48325f0e20>), LightRAGRetrieverPipeline(DS=<theflow.base.unset_ object at 0x7f48325f0e20>, FSPath=<theflow.base.unset_ object at 0x7f48325f0e20>, Index=<class 'ktem.index.file.index.IndexTable'>, Source=<theflow.base.unset_ object at 0x7f48325f0e20>, VS=<theflow.base.unset_ object at 0x7f48325f0e20>, file_ids=['219b3c04-784d-4c2f-b353-ef6610ff9e82'], search_type='local', user_id=<theflow.base.unset_ object at 0x7f48325f0e20>)]

Browsers

No response

OS

Linux

Additional information

No response

[Meteor-de]

大致如此，三个文档中，1、3文档可以在图中显示，但是2文档完全无法得到任何回复。同时，如果在单选文档1或3时，都会得到一个二者混合的图和回答，并不能有效的选择其中一个。 To be more precise, among the three documents I tested, Document 1 and Document 3 can be partially visualized in the graph. However, Document 2 fails to elicit any response or graph whatsoever. At the same time, if I try to select only Document 1 or only Document 3 for a chat, the system provides a hybrid graph and a mixed answer that combines information from both documents. It does not effectively isolate the conversation to the single selected document.

[Meteor-de]

To create a public link, set share=True in launch().
User-id: None, can see public conversations: False
User-id: 92021f5bde6949cca8dc7a6b2e983844, can see public conversations: True
User-id: 92021f5bde6949cca8dc7a6b2e983844, can see public conversations: True
Session reasoning type None use mindmap True use citation highlight language en
Session LLM
Reasoning class <class 'ktem.reasoning.simple.FullQAPipeline'>
Reasoning state {'app': {'regen': False}, 'pipeline': {}}
Thinking ...
Retrievers [DocumentRetrievalPipeline(DS=<kotaemon.storages.docstores.lancedb.LanceDBDocumentStore object at 0x7febaa6d27a0>, FSPath=PosixPath('/root/kotaemon/ktem_app_data/user_data/files/index_1'), Index=<class 'ktem.index.file.index.IndexTable'>, Source=<class 'ktem.index.file.index.Source'>, VS=<kotaemon.storages.vectorstores.chroma.ChromaVectorStore object at 0x7febaa6d18d0>, get_extra_table=False, llm_scorer=LLMTrulensScoring(concurrent=True, normalize=10, prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7fec094ff8b0>, system_prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7fec094ff880>, top_k=3, user_prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7febb0ed88b0>), mmr=False, rerankers=[CohereReranking(base_url=None, cohere_api_key='<COHERE_API_KEY>', model_name='rerank-multilingual-v2.0')], retrieval_mode='hybrid', top_k=10, user_id='92021f5bde6949cca8dc7a6b2e983844'), GraphRAGRetrieverPipeline(DS=<theflow.base.unset_ object at 0x7fec6cad0e80>, FSPath=<theflow.base.unset_ object at 0x7fec6cad0e80>, Index=<class 'ktem.index.file.index.IndexTable'>, Source=<theflow.base.unset_ object at 0x7fec6cad0e80>, VS=<theflow.base.unset_ object at 0x7fec6cad0e80>, file_ids=[], user_id=<theflow.base.unset_ object at 0x7fec6cad0e80>), LightRAGRetrieverPipeline(DS=<theflow.base.unset_ object at 0x7fec6cad0e80>, FSPath=<theflow.base.unset_ object at 0x7fec6cad0e80>, Index=<class 'ktem.index.file.index.IndexTable'>, Source=<theflow.base.unset_ object at 0x7fec6cad0e80>, VS=<theflow.base.unset_ object at 0x7fec6cad0e80>, file_ids=['b427879a-453f-4c6c-82fa-3f57f17291a0'], search_type='local', user_id=<theflow.base.unset_ object at 0x7fec6cad0e80>)]
searching in doc_ids []
GraphRAG embedding dim 768
INFO:nano-vectordb:Load (228, 768) data
INFO:nano-vectordb:Init {'embedding_dim': 768, 'metric': 'cosine', 'storage_file': '/root/kotaemon/ktem_app_data/user_data/files/lightrag/9209aaa5-0be8-4f08-9ec8-4678cdbccc6a/input/vdb_entities.json'} 228 data
INFO:nano-vectordb:Load (159, 768) data
INFO:nano-vectordb:Init {'embedding_dim': 768, 'metric': 'cosine', 'storage_file': '/root/kotaemon/ktem_app_data/user_data/files/lightrag/9209aaa5-0be8-4f08-9ec8-4678cdbccc6a/input/vdb_relationships.json'} 159 data
INFO:nano-vectordb:Load (6, 768) data
INFO:nano-vectordb:Init {'embedding_dim': 768, 'metric': 'cosine', 'storage_file': '/root/kotaemon/ktem_app_data/user_data/files/lightrag/9209aaa5-0be8-4f08-9ec8-4678cdbccc6a/input/vdb_chunks.json'} 6 data
Rerank is enabled but no rerank_model_func provided. Reranking will be skipped.
search_type local
INFO:root:Local query uses 40 entities, 39 relations, 2 text units
pdf-path: does not exist
pdf-path: does not exist
pdf-path: does not exist
Got 3 retrieved documents
len (original) 17449
len (trimmed) 17449
Got 0 images
Trying LLM streaming
CitationPipeline: invoking LLM
LLM rerank scores [0.5, 0.5, 0.5]
Got 0 cited docs
pdf-path: does not exist
pdf-path: does not exist
pdf-path: does not exist
Session reasoning type None use mindmap True use citation highlight language en
Session LLM
Reasoning class <class 'ktem.reasoning.simple.FullQAPipeline'>
Reasoning state {'app': {'regen': False}, 'pipeline': {}}
Thinking ...
Retrievers [DocumentRetrievalPipeline(DS=<kotaemon.storages.docstores.lancedb.LanceDBDocumentStore object at 0x7febaa6d27a0>, FSPath=PosixPath('/root/kotaemon/ktem_app_data/user_data/files/index_1'), Index=<class 'ktem.index.file.index.IndexTable'>, Source=<class 'ktem.index.file.index.Source'>, VS=<kotaemon.storages.vectorstores.chroma.ChromaVectorStore object at 0x7febaa6d18d0>, get_extra_table=False, llm_scorer=LLMTrulensScoring(concurrent=True, normalize=10, prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7feb32aaef50>, system_prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7feb32aaf3a0>, top_k=3, user_prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7feb32aae800>), mmr=False, rerankers=[CohereReranking(base_url=None, cohere_api_key='<COHERE_API_KEY>', model_name='rerank-multilingual-v2.0')], retrieval_mode='hybrid', top_k=10, user_id='92021f5bde6949cca8dc7a6b2e983844'), GraphRAGRetrieverPipeline(DS=<theflow.base.unset_ object at 0x7fec6cad0e80>, FSPath=<theflow.base.unset_ object at 0x7fec6cad0e80>, Index=<class 'ktem.index.file.index.IndexTable'>, Source=<theflow.base.unset_ object at 0x7fec6cad0e80>, VS=<theflow.base.unset_ object at 0x7fec6cad0e80>, file_ids=[], user_id=<theflow.base.unset_ object at 0x7fec6cad0e80>), LightRAGRetrieverPipeline(DS=<theflow.base.unset_ object at 0x7fec6cad0e80>, FSPath=<theflow.base.unset_ object at 0x7fec6cad0e80>, Index=<class 'ktem.index.file.index.IndexTable'>, Source=<theflow.base.unset_ object at 0x7fec6cad0e80>, VS=<theflow.base.unset_ object at 0x7fec6cad0e80>, file_ids=['fcb78931-bbef-4494-be77-a6a196dcc5dd'], search_type='local', user_id=<theflow.base.unset_ object at 0x7fec6cad0e80>)]
searching in doc_ids []
INFO:ktem.index.file.pipelines:Skip retrieval because of no selected files: DocumentRetrievalPipeline(
(vector_retrieval): <function Function.prepare_child..exec at 0x7feb400dbc70>
(embedding): <function Function.prepare_child..exec at 0x7feb400dbeb0>
)
GraphRAG embedding dim 768
INFO:nano-vectordb:Load (228, 768) data
INFO:nano-vectordb:Init {'embedding_dim': 768, 'metric': 'cosine', 'storage_file': '/root/kotaemon/ktem_app_data/user_data/files/lightrag/9209aaa5-0be8-4f08-9ec8-4678cdbccc6a/input/vdb_entities.json'} 228 data
INFO:nano-vectordb:Load (159, 768) data
INFO:nano-vectordb:Init {'embedding_dim': 768, 'metric': 'cosine', 'storage_file': '/root/kotaemon/ktem_app_data/user_data/files/lightrag/9209aaa5-0be8-4f08-9ec8-4678cdbccc6a/input/vdb_relationships.json'} 159 data
INFO:nano-vectordb:Load (6, 768) data
INFO:nano-vectordb:Init {'embedding_dim': 768, 'metric': 'cosine', 'storage_file': '/root/kotaemon/ktem_app_data/user_data/files/lightrag/9209aaa5-0be8-4f08-9ec8-4678cdbccc6a/input/vdb_chunks.json'} 6 data
Rerank is enabled but no rerank_model_func provided. Reranking will be skipped.
search_type local
INFO:root:Local query uses 40 entities, 41 relations, 3 text units
pdf-path: does not exist
pdf-path: does not exist
pdf-path: does not exist
Got 3 retrieved documents
len (original) 18081
len (trimmed) 18081
Got 0 images
Trying LLM streaming
CitationPipeline: invoking LLM
CitationPipeline: finish invoking LLM
CitationPipeline: {"evidences":["里面使用的模型是基于文本生成和理解的模型，用于回答问题和提供相关信息。"]}
User-id: 92021f5bde6949cca8dc7a6b2e983844, can see public conversations: True
LLM rerank scores [0.4, 0.2, 0.2]
Got 0 cited docs
pdf-path: does not exist
pdf-path: does not exist
pdf-path: does not exist
CitationPipeline: finish invoking LLM
Session reasoning type None use mindmap True use citation highlight language en
Session LLM
Reasoning class <class 'ktem.reasoning.simple.FullQAPipeline'>
Reasoning state {'app': {'regen': False}, 'pipeline': {}}
Thinking ...
Retrievers [DocumentRetrievalPipeline(DS=<kotaemon.storages.docstores.lancedb.LanceDBDocumentStore object at 0x7febaa6d27a0>, FSPath=PosixPath('/root/kotaemon/ktem_app_data/user_data/files/index_1'), Index=<class 'ktem.index.file.index.IndexTable'>, Source=<class 'ktem.index.file.index.Source'>, VS=<kotaemon.storages.vectorstores.chroma.ChromaVectorStore object at 0x7febaa6d18d0>, get_extra_table=False, llm_scorer=LLMTrulensScoring(concurrent=True, normalize=10, prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7fec094fe5f0>, system_prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7feb32b20640>, top_k=3, user_prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7feb32951d20>), mmr=False, rerankers=[CohereReranking(base_url=None, cohere_api_key='<COHERE_API_KEY>', model_name='rerank-multilingual-v2.0')], retrieval_mode='hybrid', top_k=10, user_id='92021f5bde6949cca8dc7a6b2e983844'), GraphRAGRetrieverPipeline(DS=<theflow.base.unset object at 0x7fec6cad0e80>, FSPath=<theflow.base.unset object at 0x7fec6cad0e80>, Index=<class 'ktem.index.file.index.IndexTable'>, Source=<theflow.base.unset_ object at 0x7fec6cad0e80>, VS=<theflow.base.unset_ object at 0x7fec6cad0e80>, file_ids=[], user_id=<theflow.base.unset_ object at 0x7fec6cad0e80>), LightRAGRetrieverPipeline(DS=<theflow.base.unset_ object at 0x7fec6cad0e80>, FSPath=<theflow.base.unset_ object at 0x7fec6cad0e80>, Index=<class 'ktem.index.file.index.IndexTable'>, Source=<theflow.base.unset_ object at 0x7fec6cad0e80>, VS=<theflow.base.unset_ object at 0x7fec6cad0e80>, file_ids=['46643095-33ea-423d-959b-22453b49a0ae'], search_type='local', user_id=<theflow.base.unset_ object at 0x7fec6cad0e80>)]
searching in doc_ids []
INFO:ktem.index.file.pipelines:Skip retrieval because of no selected files: DocumentRetrievalPipeline(
(vector_retrieval): <function Function.prepare_child..exec at 0x7feb883c48b0>
(embedding): <function Function.prepare_child..exec at 0x7feb883c4940>
)
GraphRAG embedding dim 768
INFO:nano-vectordb:Load (228, 768) data
INFO:nano-vectordb:Init {'embedding_dim': 768, 'metric': 'cosine', 'storage_file': '/root/kotaemon/ktem_app_data/user_data/files/lightrag/9209aaa5-0be8-4f08-9ec8-4678cdbccc6a/input/vdb_entities.json'} 228 data
INFO:nano-vectordb:Load (159, 768) data
INFO:nano-vectordb:Init {'embedding_dim': 768, 'metric': 'cosine', 'storage_file': '/root/kotaemon/ktem_app_data/user_data/files/lightrag/9209aaa5-0be8-4f08-9ec8-4678cdbccc6a/input/vdb_relationships.json'} 159 data
INFO:nano-vectordb:Load (6, 768) data
INFO:nano-vectordb:Init {'embedding_dim': 768, 'metric': 'cosine', 'storage_file': '/root/kotaemon/ktem_app_data/user_data/files/lightrag/9209aaa5-0be8-4f08-9ec8-4678cdbccc6a/input/vdb_chunks.json'} 6 data
Rerank is enabled but no rerank_model_func provided. Reranking will be skipped.
search_type local
INFO:root:Local query uses 40 entities, 37 relations, 3 text units
pdf-path: does not exist
pdf-path: does not exist
pdf-path: does not exist
Got 3 retrieved documents
len (original) 17897
len (trimmed) 17897
Got 0 images
Trying LLM streaming
CitationPipeline: invoking LLM
LLM rerank scores [0.2, 0.2, 0.2]
CitationPipeline: finish invoking LLM
CitationPipeline: {"evidences":["分享","我是从天空俯视着地面is an event where the player describes themselves as looking down from the sky, indicating a unique perspective or position.","我来自那里is an event where the player describes their origin, possibly indicating a mysterious or significant background.","我是从深海仰望着天空is an event where the player describes themselves as looking up from the deep sea, indicating introspection or a unique identity.","天空是指一个广阔的空间，与深海形成对比，象征着开放或可见的领域。"]}
pdf-path: does not exist
Got 1 cited docs
pdf-path: does not exist
pdf-path: does not exist
Session reasoning type None use mindmap True use citation highlight language en
Session LLM
Reasoning class <class 'ktem.reasoning.simple.FullQAPipeline'>
Reasoning state {'app': {'regen': False}, 'pipeline': {}}
Thinking ...
Retrievers [DocumentRetrievalPipeline(DS=<kotaemon.storages.docstores.lancedb.LanceDBDocumentStore object at 0x7febaa6d27a0>, FSPath=PosixPath('/root/kotaemon/ktem_app_data/user_data/files/index_1'), Index=<class 'ktem.index.file.index.IndexTable'>, Source=<class 'ktem.index.file.index.Source'>, VS=<kotaemon.storages.vectorstores.chroma.ChromaVectorStore object at 0x7febaa6d18d0>, get_extra_table=False, llm_scorer=LLMTrulensScoring(concurrent=True, normalize=10, prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7febb027e200>, system_prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7feb32ac1990>, top_k=3, user_prompt_template=<kotaemon.llms.prompts.template.PromptTemplate object at 0x7feb32ac3430>), mmr=False, rerankers=[CohereReranking(base_url=None, cohere_api_key='<COHERE_API_KEY>', model_name='rerank-multilingual-v2.0')], retrieval_mode='hybrid', top_k=10, user_id='92021f5bde6949cca8dc7a6b2e983844'), GraphRAGRetrieverPipeline(DS=<theflow.base.unset object at 0x7fec6cad0e80>, FSPath=<theflow.base.unset object at 0x7fec6cad0e80>, Index=<class 'ktem.index.file.index.IndexTable'>, Source=<theflow.base.unset_ object at 0x7fec6cad0e80>, VS=<theflow.base.unset_ object at 0x7fec6cad0e80>, file_ids=[], user_id=<theflow.base.unset_ object at 0x7fec6cad0e80>), LightRAGRetrieverPipeline(DS=<theflow.base.unset_ object at 0x7fec6cad0e80>, FSPath=<theflow.base.unset_ object at 0x7fec6cad0e80>, Index=<class 'ktem.index.file.index.IndexTable'>, Source=<theflow.base.unset_ object at 0x7fec6cad0e80>, VS=<theflow.base.unset_ object at 0x7fec6cad0e80>, file_ids=['46643095-33ea-423d-959b-22453b49a0ae'], search_type='local', user_id=<theflow.base.unset_ object at 0x7fec6cad0e80>)]
searching in doc_ids []
INFO:ktem.index.file.pipelines:Skip retrieval because of no selected files: DocumentRetrievalPipeline(
(vector_retrieval): <function Function._prepare_child..exec at 0x7febabdab760>
(embedding): <function Function._prepare_child..exec at 0x7febabda9990>
)
GraphRAG embedding dim 768
INFO:nano-vectordb:Load (228, 768) data
INFO:nano-vectordb:Init {'embedding_dim': 768, 'metric': 'cosine', 'storage_file': '/root/kotaemon/ktem_app_data/user_data/files/lightrag/9209aaa5-0be8-4f08-9ec8-4678cdbccc6a/input/vdb_entities.json'} 228 data
INFO:nano-vectordb:Load (159, 768) data
INFO:nano-vectordb:Init {'embedding_dim': 768, 'metric': 'cosine', 'storage_file': '/root/kotaemon/ktem_app_data/user_data/files/lightrag/9209aaa5-0be8-4f08-9ec8-4678cdbccc6a/input/vdb_relationships.json'} 159 data
INFO:nano-vectordb:Load (6, 768) data
INFO:nano-vectordb:Init {'embedding_dim': 768, 'metric': 'cosine', 'storage_file': '/root/kotaemon/ktem_app_data/user_data/files/lightrag/9209aaa5-0be8-4f08-9ec8-4678cdbccc6a/input/vdb_chunks.json'} 6 data
Rerank is enabled but no rerank_model_func provided. Reranking will be skipped.
search_type local
INFO:root:Local query uses 40 entities, 33 relations, 3 text units
pdf-path: does not exist
pdf-path: does not exist
pdf-path: does not exist
Got 3 retrieved documents
len (original) 16677
len (trimmed) 16677
Got 0 images
Trying LLM streaming
CitationPipeline: invoking LLM
LLM rerank scores [0.2, 0.2, 0.2]
CitationPipeline: finish invoking LLM
Got 0 cited docs
pdf-path: does not exist
pdf-path: does not exist
pdf-path: does not exist
这里是完整的报错日志，他正确的改变文件id，但是调用的数据库却是有问题的，仍然是之前的。
Here is the complete log trace.
The logs confirm that the application correctly changes the file_id in the LightRAGRetrieverPipeline each time I select a new document for a new chat session.
However, the core problem is that the application fails to load the corresponding knowledge base. Instead, it consistently loads the database from a previous, incorrect document. This behavior persists across all sessions.

[onestardao]

It looks like the issue comes from session state not being properly reset when switching documents.
From your logs and description:

The retriever still binds to the first uploaded document (Doc A) even after selecting a new one (Doc B), so the graph and answers remain tied to the earlier session context.

When you select only Doc 1 or Doc 3, the system is generating a hybrid graph that mixes nodes from multiple documents instead of isolating the selected one.

This indicates that the LightRAGRetrieverPipeline is not clearing or re-initializing the active knowledge graph state when a fresh session starts.

So the root problem isn’t ingestion itself (your docs are indexed fine), but that the per-session file_id / graph binding is persisting across sessions. This is why Document 2 fails to trigger responses at all, while Documents 1 and 3 “bleed” into each other.

You might want to check whether session.reset() or equivalent graph clearing is being called when you start a new chat with a different document.

[Meteor-de]

It looks like the issue comes from session state not being properly reset when switching documents. From your logs and description:

The retriever still binds to the first uploaded document (Doc A) even after selecting a new one (Doc B), so the graph and answers remain tied to the earlier session context.

When you select only Doc 1 or Doc 3, the system is generating a hybrid graph that mixes nodes from multiple documents instead of isolating the selected one.

This indicates that the LightRAGRetrieverPipeline is not clearing or re-initializing the active knowledge graph state when a fresh session starts.

So the root problem isn’t ingestion itself (your docs are indexed fine), but that the per-session file_id / graph binding is persisting across sessions. This is why Document 2 fails to trigger responses at all, while Documents 1 and 3 “bleed” into each other.

You might want to check whether session.reset() or equivalent graph clearing is being called when you start a new chat with a different document.

Yes, the function does not include a graph clearing operation. I believe this is likely the cause of the document graph errors. How should I resolve this issue?

[onestardao]

Thanks for confirming — that makes sense, since the retriever session state not being properly reset explains why the hybrid graph keeps leaking nodes from the previous document.

This type of issue (session persistence vs. document scoping) is one of the recurring patterns I’ve seen when testing RAG/graph retrieval systems. To avoid repeating the same debugging cycle, I maintain a public checklist of failure modes and their root causes (MIT-licensed):

👉 Workflow Failure & Graph Retrieval Patterns

You’ll find “session not reset → cross-document contamination” mapped there, along with other cases like partial ingestion, scope drift, and cache leakage. Might help you compare with your setup and spot if the fix is only about clearing graph state, or if other persistence layers (session store, cache binding) also need reset.