mirror of
https://github.com/Mintplex-Labs/anything-llm.git
synced 2026-07-19 14:13:35 -04:00
[GH-ISSUE #5533] Bug: Changing DeepSeek model or API key resets conversation context (workspace forgets history) #5150
Closed
opened 2026-06-05 14:52:19 -04:00 by yindo
·
4 comments
No Branch/Tag Specified
master
5846-bug-when-scrolling-up-scrolling-jumps
refactor-remove-workspace-pfp
5969-bug-stopgenerationbutton-disappears-on-the-first-prompt-that-initiates-an-agent-session
2235-bug-how-to-upload-a-folder-with-subfolders-with-files-to-anythingllm
5990-workspace-update-fails-with-unknown-argument-router_id-v1130v1150-intel-mac
opencomputer-examples
pg
feat/image-generation-translations
feat/image-generation
5924-bug-meeting-summary-fails-with-sincludes-is-not-a-function-when-default-llm-is-anthropic-claude
render
feat/uniform-modal-component
5883-bug-unescaped-content-in-json-strings-being-passed-to-document-generator-tools
5901-bug-api-update-embeddings-fails-prisma-argument-filename-is-missing-on-workspace_documentscreate-desktop-windows-v1141
hybrid-search
1981-translations
5752-bug-prompts-to-local-jan-endpoint-unresponsive
feat-disable-native-tool-calling-env-var
5676-bug-non-ollama-agent-providers-do-not-parse-and-present-reasoning-content
feat/markdown-web-scraping
5717-bug-apiv1documentupload-silently-drops-metadata-field-in-desktop-1130-arg-count-mismatch-nested-payload-key
fix/aibitat-context-overflow
5711-bug-erratic-deepseek-v4-flash-the-agent-model-failed-to-respond-400-the-reasoning_content
5631-feat-custom-api-request-timeouts-for-ai-providers
feat-reasoning-control
feat-agent-clarifying-questions-translations
5583-bug-lm-studio-provider-does-not-present-reasoning-output
5313-normalize-translations
feat/memory-translations
5305-lemonade-embedding-engine-swallows-errors-falsely-reports-documents-as-embedded
5060-bug-agent-interactions-agent-are-not-persisted-to-thread-history-via-api
pptx-subagent
feat-render-images-from-mcp-tool-results
stt-provider-expansion-openai-api-compatible
feat-file-search-agent-tool
feat-native-embedder-job-queue
5189-normalize-translations
feat-file-search-agent-tool-translations
i18n-eslint
5140-auto-migration
5112-bug-openrouter-failed-message-bug
3506-feat-parameters-for-openrouter-models
4992-feat-preserve-scroll-position
4973-bug-markdown-numbered-list-display-in-reasoning-pane
desktop
4938-bug-pending-chat-rerendering-ui-bug
quickstart-env
node-llama-cpp-in-container-cuda
node-llama-cpp-in-container
ollama-in-container
4817-feat-set-cooldown-per-mcp-server
4845-keyboard-shortcuts-to-navigate-in-chat
4844-feat-reorder-threads-by-latest-interaction
standardize-username-constraints-normalize-translations
4792-feat-refactor-workspacepfp-image
1382-embed-ip-improvements
1382-bug-embed-api-improvements
refactor-eslint-frontend
4687-feat-refactor-vector-db-providers
4615-feat-disable-apidocs-with-environment-variable
4559-feat-agent-web-search-enable-ordering-of-results
4599-bug-ollama-race-condition-bug
4572-bug-lmstudio-provided-llm-stopped-working-with-anything-llm-after-upgrading-to-190
4508-agent-youtube-transcript-analysis
4497-feat-workspace-names
frontend-eslint
ollama-lmstudio-auto-context-window
4431-validate-vector-database-connectioN
2019-slash-command-keyboard-selection
microsoft-foundry-provider
4431-validate-vector-database-connection
4325-sys-prompt-var-improvements
3209-feat-apiv1workspacestream-chat-sources-citations
4210-bug-voice-to-text-overwrite
4136-feat-jan-as-a-backend-server-option
4172-feat-openai-o3-support
1.8.3-rerelease
web-push-notifications-service
tasks
3955-feat-jinaai-embedder-provider-support
3921-feat-agent-skills-uiux-improvements
3901-bug-validfunccall-checks-optional-arguments
keyboard-dev
1787-custom-roles-and-permissions
add-jira-slack-data-connector
office-extension-wip
lightmode-dropdown-color-update
3586-bug-agent-flow-function-description-provided-by-user-is-not-seen-in-the-llm-query
3463-bug-agent-continues-to-run-if-request-failed-even-after-exit
3439-feat-call-variables-within-the-flow-api-block-url-field
3282-manager-view-models-workspace
3280-token-counting-server-side-truncation-improvements
3147-bug-embedded-chat-widget---not-considering-query-mode-option-always-working-in-chat-mode
2995-feat-disable-temperature-setting-for-deepseek-r1-deepseek-reasoner-model
2827-feat-perplexity-citations
2866-feat-finally-a-gemini-models-endpoint
2647-feat-hpp-header-for-a-c++-code-file-mime-addition
lancedb-revert
1656-feat-implement-tooltip-ui-designs
2011-feat-bump-perplexity-models
1873-feat-auto-add-and-watch-folder-for-document-uploads
1297-feat-gemini-agent-support
1759-bug-ui-bug-fixes
1686-feat-implement-winston-for-logging
1536-bug-toggling-on-users-can-delete-workspaces-does-not-take-effect
agent-ui-mobile-styles
1522-feat-chromadb-support
1595-bug-unable-to-get-live-web-search-and-browsing-agent-working-using-google-custom-search-engine-error-getaddrinfo-enotfound-http-errno-3008
1582-bug-lm-studio-does-not-allow-for-different-model-selection
1312-bug-usernames-should-not-be-case-sensitive-when-logging-in
1029-feat-hf-serverless-inference-api
1086-feat-implement-normalized-input-fields
knowledge-graph-support
644-bug-uploaded-file-name-does-not-match-the-displayed-file-name-after-the-upload
v1.15.0
v1.14.2
v1.14.1
v1.14.0
v1.13.0
v1.12.1
v1.12.0
v1.11.2
v1.11.1
v1.11.0
v1.10.0
v1.9.1
v1.9.0
v1.8.5
v1.8.4
v1.8.3
v1.8.2
v1.8.1
v1.8.0
v1.7.8
v1.7.6
v1.7.5
v1.7.4
v1.4.0
v1.3.0
v1.2.4
v1.2.3
v1.2.2
v1.2.1
v1.2.0
v1.1.1
v1.1.0
v1.0.0
Labels
Clear labels
Desktop
Docker
Integration Request
Integration Request
OS: Linux
OS: Mobile
OS: Windows
UI/UX
blocked
bug
bug
core-team-only
documentation
duplicate
embed-widget
enhancement
feature request
github_actions
good first issue
investigating
needs info / can't replicate
possible bug
pull-request
question
stage: specifications
wontfix
Mirrored from GitHub Pull Request
Milestone
No items
No Milestone
Projects
Clear projects
No project
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: Mintplex-Labs/anything-llm#5150
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Delete Branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @miraitech24 on GitHub (Apr 25, 2026).
Original GitHub issue: https://github.com/Mintplex-Labs/anything-llm/issues/5533
Problem
When I change the DeepSeek model (e.g., from deepseek-chat to deepseek-v4-flash) or regenerate the API key, the workspace treats every message as a first-time interaction. It does not remember any previous conversation history.
What happens
What should happen
Changing API keys or model names should NOT break conversation continuity. The workspace should maintain its history and continue sending full context to the new API key/model.
Root cause suspicion
AnythingLLM seems to treat model/key changes as a 'new session' instead of preserving the existing conversation thread. The UI shows old messages, but they are not being included in the API request payload.
Steps to reproduce
Expected vs actual
Why this matters
LLM APIs are stateless by design — that's correct. The application's job is to manage state. If AnythingLLM fails to send conversation history after settings changes, it's not fulfilling its core purpose as a chat/workspace application.
Additional note
I confirmed the API key is valid using curl:
Authentication works. The issue is on AnythingLLM's side, not DeepSeek's.
Environment
@timothycarambat commented on GitHub (Apr 25, 2026):
This is almost certainly a model thing - we are most certainly send the message history with every provider (not just deepseek). That part is normalized across all providers. How are you validating that it does not "remember" anything?
The model could just be that bad at recall - its not impossible!
@miraitech24 commented on GitHub (Apr 27, 2026):
Regarding your statement: 'Pinning a document = always viewable like a bulletin board'
Thank you for your response. However, I believe there is a critical misunderstanding here about what 'always viewable' means, which has now been confirmed by my actual usage.
What I expected (and the natural interpretation of 'bulletin board'):
→ This does NOT happen.
What your documentation and source code say 'Pinning' actually does:
Crucially, there is NO mention of cross-session or cross-thread persistence. The implementation detail shows
pinnedWorkspacesties the pin ONLY to the current workspace context.In simple terms:
These are completely different features. If my interpretation is wrong, and the feature is actually supposed to work like a bulletin board, then your documentation is incorrect and there is a significant bug in the implementation (likely the same issue reported in #2841, where the API disregards pinned documents).
Could you please clarify:
Thank you.
@miraitech24 commented on GitHub (Apr 28, 2026):
根本的な設計思想の欠陥について
度重なる議論、ご対応ありがとうございます。しかし、今回の一連の問題は個別のバグではなく、AnythingLLMの『設計思想の欠陥』に起因している可能性が高いという結論に至りました。
問題の本質:
AnythingLLMは、LLMとユーザーの間に『仲介役』として入り、『ワークスペース』『埋め込み』『ピン留め』といった独自の概念を提供しています。しかし、これらの概念はLLMに伝わるものではなく、単なるローカルの管理タグです。
ユーザーは『ワークスペースにファイルを埋め込んだ』『ピン留めした』という操作をしますが、LLMに伝わるのは『あるテキスト断片がプロンプトに追加された』という事実だけです。その結果、『埋め込みを見える?』『ピン留めしたファイルの内容は?』といったユーザーの問いかけに対して、LLMは『ファイルシステムにアクセスできない』と混乱した回答を返します。
これはバグではなく、設計思想の欠陥です。 なぜなら、システムは設計通りに動作している可能性があるからです。問題は、その『設計』自体がユーザーの期待と現実のLLMの動作を著しく乖離させている点にあります。
提案:
これは技術的に難しい問題ですが、少なくとも以下の2つの方向性で改善できないでしょうか。
RAGの本質を変えろとは言いません。しかし、ユーザーが『埋め込み』という機能名から自然に抱く期待と、実際の動作のギャップは、現在大きすぎます。これは製品の競争力を大きく損なう要因です。
長期的な視点での設計の見直しを、ぜひご検討ください。
@miraitech24 commented on GitHub (Apr 28, 2026):
検証完了:ピン留めされたファイルの中身はモデルに渡されていない
以下の検証を実施しました。
検証手順:
結果:
モデルが表示したテキストには、以下のものが含まれていました。
しかし、以下のものは含まれていませんでした。
結論:
AnythingLLMは『ピン留め』されたファイルの中身をモデルに渡していません。『ファイル名』は渡しているため、モデルは『BIOCore.csv』という名前は認識できるが、その内容を知らない、という状態です。
これはバグではなく、設計思想の欠陥です。なぜなら、システムはおそらく『ピン留めされたファイルの中身は渡さない』という設計だからです。
しかし、ユーザーは『ピン留め』という言葉から『そのファイルを永続的にコンテキストに挿入する』と期待します。その期待と実際の動作の乖離が、今回の混乱のすべての原因です。
改善提案:
この検証結果をもって、本Issueを『設計思想の欠陥』としてクローズします。今後の設計の参考にしていただければ幸いです。