- Implemented a check to reject file paths containing "..", enhancing security against path traversal vulnerabilities in the file serving functionality.
- Added SSRF protection by stripping client-supplied URL and Caption fields from image attachments in the QA request handler.
- Introduced a validation function to ensure Feishu API path parameters contain only safe characters, preventing path traversal attacks.
- Enhanced the WeCom webhook adapter to reject internal/private URLs unless they are on an allowlist, improving security during file downloads.
- Implemented a mechanism to bypass SSRF checks for trusted IM platform API hosts.
- Updated the assistant descriptions in various prompt templates to specify that they are developed by Tencent.
- Added new intent-specific system prompt templates for greeting, chitchat, follow-up, image analysis, summarization, and web search unavailability scenarios.
- Enhanced citation rules in the system prompts to ensure proper inline citation formatting.
- Refactored the rewrite prompt to classify user intent more accurately and output structured JSON responses.
- Changed the kill delay from 30 seconds to 2 seconds in the configuration.
- Refactored the server startup process to include a retry mechanism for binding to the port, addressing potential issues during hot-reload scenarios.
- Added a new function, listenWithRetry, to handle port binding with exponential backoff, improving server reliability during restarts.
- Added token usage tracking to the AgentEngine, allowing for better estimation of current context token counts.
- Implemented a new method to estimate current tokens based on previous usage and newly appended messages, improving context management.
- Updated context window management to utilize estimated token counts for more efficient message consolidation.
- Enhanced streaming methods to include token usage information, providing better insights into LLM interactions.
- Refactored the token estimator to support BPE tokenization, ensuring accurate token counts for messages.
- Added unit tests to validate new token estimation and context management functionalities.
- Added functionality to resolve embedded data:image;base64 images in markdown, storing them and replacing references with stable URLs.
- Updated the knowledge service to process data URIs before handling remote images, ensuring all images are resolved consistently.
- Improved logging for resolved images, providing better insights during manual processing.
- Added unit tests to validate the new data URI resolution functionality and ensure reliability.
- Implemented a new method in the chunk repository to find FAQ chunks with overlapping standard or similar questions.
- Updated the knowledge service to utilize this method for checking duplicates during FAQ question creation.
- Improved error handling for duplicate questions, providing clearer feedback on conflicts.
- Added a new method to render graph extraction prompts with shared placeholders for language.
- Updated entity and relationship extraction methods to utilize the new rendering function, improving prompt customization.
- Enhanced question generation task to include language context, ensuring prompts are generated in the correct language.
- Improved error handling for empty prompt configurations in question generation, enhancing robustness.
- Removed the query display section from the DatabaseQuery component to streamline the UI.
- Updated the DatabaseQueryData interface to eliminate the query field, reflecting changes in tool result handling.
- Enhanced the DatabaseQuery tool execution to focus on returning structured results without exposing raw SQL queries.
- Added a new utility to strip <think> blocks from LLM outputs, improving content clarity and user experience.
- Introduced the AgentEngine with methods for executing tool calls, managing context windows, and handling final answer generation.
- Added functionality to stream LLM responses and tool results through the EventBus, enhancing real-time interaction.
- Implemented context management to consolidate messages and prevent exceeding token limits, ensuring efficient use of context.
- Enhanced error handling and logging for tool execution and response analysis, improving robustness and traceability.
- Added unit tests to validate the new functionalities and ensure reliability in agent operations.
- Introduced a default timeout for tool execution to prevent indefinite hangs during long-running tasks.
- Implemented context cancellation handling in the execution loop to gracefully manage timeouts and user cancellations.
- Added a JSON repair utility to fix common malformations in tool arguments, improving robustness in argument parsing.
- Enhanced message sanitization to merge consecutive user messages and handle orphaned tool results, ensuring compatibility with LLM providers.
- Added unit tests for the new JSON repair and message sanitization functionalities to validate their effectiveness.
- Introduced a memory consolidator to summarize older messages when the context window exceeds a specified limit, preserving key facts and tool results.
- Enhanced the AgentEngine to utilize the new token estimator and memory consolidator for improved context window management.
- Implemented logic to compress message history when approaching token limits, ensuring efficient use of context.
- Added unit tests for the memory consolidator and token estimator to validate functionality and performance.
- Updated tool execution to validate parameters against JSON schema before execution, improving error handling and robustness.
- Added a new `ProviderSpec` struct to define provider-specific behaviors for chat completions.
- Implemented request and endpoint customizers for various providers including Qwen3, LKEAP, DeepSeek, and Generic.
- Refactored the `NewRemoteChat` function to utilize the new provider specs for customizing requests based on the provider and model name.
- Removed legacy chat implementations for LKEAP, NVIDIA, DeepSeek, and Qwen, consolidating their logic into the new structure.
- Introduced a configurable LLM call timeout with a default value, allowing for better control over LLM call durations.
- Added logic to retry transient errors (e.g., timeouts, rate limits) up to a specified maximum number of retries, improving robustness in error handling.
- Implemented parameter casting for tool arguments to ensure correct types are used, addressing common LLM quirks.
- Enhanced tool execution error messages with guidance for retrying with different approaches.
- Added validation for configuration values to prevent runtime errors.
Previously, rag_stream relied entirely on the rewrite plugin's
internal loadHistory() to load conversation history. This meant that
when rewrite was disabled (enable_rewrite=false), RAG mode had NO
conversation history at all — multi-turn conversations silently
degraded to single-turn.
Changes:
1. Add LOAD_HISTORY as the first step in rag_stream pipeline,
ensuring history is always available regardless of rewrite config.
2. Make rewrite plugin reuse chatManage.History if already populated
by the upstream LOAD_HISTORY step, avoiding duplicate DB queries.
This brings rag_stream in line with chat_history_stream which already
has LOAD_HISTORY as an independent step.
Replace v-html + DOMPurify sanitization with plain Vue text
interpolation ({{ }}) and CSS white-space: pre-wrap for the
thinking content display.
The thinking content is raw model output that should never be
interpreted as HTML. Using v-html here created an unnecessary XSS
attack surface — even with DOMPurify, the \n→<br/> replacement
happened before sanitization, and future changes to sanitizer config
could introduce vulnerabilities.
Plain text rendering with pre-wrap:
- Eliminates XSS risk entirely (no HTML parsing)
- Preserves line breaks natively via CSS
- Removes the sanitizeHTML import and safeProcessThinkContent function
- Simpler code with identical visual output
Two consistency fixes for <think> tag parsing in index.vue:
1. Streaming path (line 579): Changed indexOf to lastIndexOf when
splitting thinking content from answer, matching the history
loading path. Prevents incorrect splitting if </think> appears
multiple times (e.g., model discusses think tags in answer).
2. History loading path (line 274-281): Removed redundant split()
call, now uses only lastIndexOf consistently. The split was
creating an unnecessary intermediate array that could produce
wrong results with multiple </think> occurrences.
Both paths now use identical logic: lastIndexOf('</think>') to find
the boundary, substring(0, index) for thinkContent, substring(index+8)
for answer content.
Two related fixes for thinking mode streaming:
1. chat_completion_stream.go: When the response channel closes
unexpectedly (context cancelled, upstream error) while thinking
is in progress, the <think> tag was left unclosed. This caused
the stored message to have an incomplete tag, making the frontend
show a perpetual "thinking..." state on history reload.
Now we emit a closing </think> tag when the channel ends with
thinkingStarted=true but thinkingEnded=false.
2. remote_api.go: The thinking Done=true event was only sent when
the first answer content chunk arrived. If the model produced
only reasoning with no answer (e.g., finish_reason set with
empty content), the thinking-done signal was never emitted.
Now we also check isDone && state.hasThinking at the end of
processStreamDelta to ensure the done event is always sent.
The assistant message content may contain <think>...</think> tags from
extended thinking mode. Previously this raw thinking content was
indexed into the chat history knowledge base, polluting embeddings
and degrading RAG retrieval quality with intermediate reasoning text.
Now we strip thinking blocks before building the passage for
IndexMessageToKB, consistent with how load_history and rewrite
plugins already strip thinking from conversation history.
Enhanced cleanPassageForRerank() to convert markdown table data rows
(| col1 | col2 |) into plain text ("col1, col2") instead of leaving
them intact. Added filtering to skip empty/whitespace-only passages
before sending to the rerank API, preventing Baidu Qianfan 400 errors
caused by empty document fragments after markdown cleaning.
When an MCP tool returns image content items, the agent engine now
automatically generates text descriptions using the configured VLM
model and appends them to the tool result message. This makes image
content accessible to LLMs that receive tool results as text.
Images are not passed directly to the LLM because Chat Completions API
does not reliably support images in tool role messages across providers
(tested: gpt-5.4 silently ignores them, Qwen3.5 processes them).
Changes:
- Add Images field to ToolResult for preserving base64 data URIs
- Add extractContentAndImages() with MIME whitelist, size/count limits
- Add ImageDescriberFunc injection to AgentEngine (no vlm package dep)
- Add describeImages() with graceful error handling and ctx cancellation
- Add VLM model injection in CreateAgentEngine (same pattern as
SetAppConfig/SetSkillsManager)
- Redact image base64 from Data map to prevent log/SSE exposure
When a tenant setting is updated via PUT /tenants/kv/{key} (e.g.,
retrieval-config, parser-engine-config), the GORM AfterFind-decrypted
plaintext api_key was silently written back to the database because
db.Updates() does not trigger the BeforeSave hook.
Guard the repository's UpdateTenant to skip writing plaintext api_key
back to the database. Pre-encrypted values (enc:v1:… from CreateTenant
/ UpdateAPIKey) are written as-is; AfterFind-decrypted plaintext is
blanked so GORM skips the column entirely.
Closes#798
- Add 26 missing i18n keys to ko-KR and ru-RU for image upload,
storage, VLM model, and multimodal features
- Add 7 new tool name keys (queryKnowledgeGraph, finalAnswer,
readSkill, executeSkillScript, dataAnalysis, dataSchema,
databaseQuery) to all 4 locales (zh-CN, en-US, ko-KR, ru-RU)
- Replace hardcoded Chinese tool name map (TOOL_NAME_DISPLAY) in
AgentStreamDisplay with i18n key references via i18n.global.t(),
merged into existing TOOL_NAME_KEYS
- Fix hardcoded Chinese error message in KnowledgeBase.vue to use
existing i18n key knowledgeBase.fileExists
Aliyun rerank API rejects empty documents with error "document index:X
should not be empty". This fix validates passage content before building
the documents array, skipping empty chunks and logging them for data
quality monitoring.
- add setup guides for both WebSocket and Webhook modes
- document streaming reply mechanisms (editMessage / AI card)
- update architecture diagrams, data model, and config reference
- Add DingTalk AI Card streaming support with create/deliver/update via OpenAPI
- Extract shared think-block rendering into reusable im.TransformThinkBlocks
- Add stream orphan reaper and edit throttling for DingTalk and Telegram
- Apply think-block transform on all reply paths (DingTalk, Telegram, Feishu)
- Add Card Template ID config to frontend IM channel panel with i18n
Add DingTalk bot adapter supporting both webhook (HmacSHA256 signature
verification) and Stream mode (via official dingtalk-stream-sdk-go),
with sessionWebhook/OpenAPI reply fallback and AccessToken caching.
Add Telegram Bot API adapter supporting both webhook and long-polling
modes with streaming replies via editMessageText, file download via
getFile API, and timing-safe secret token verification.
Add Novita AI as a new model provider supporting chat, embedding,
and VLLM model types via OpenAI-compatible API endpoint.
- Add NovitaProvider implementation with OpenAI-compatible API
- Register ProviderNovita in provider registry
- Add ModelSourceNovita constant
- Add Novita to frontend fallback provider options
- Add i18n translations for en-US, zh-CN, ru-RU, ko-KR
- Parse reasoning field together with standard stream response using local struct
- Unify reasoning content extraction from standard and vLLM response paths
- Adapt processStreamDelta to accept reasoning content as a separate parameter
- Send reasoning content as thinking response in the stream when available
- Ensure compatibility with existing SDK logic by constructing standard choice object
Reset the global page variable to 1 when reloading the file list after operations like tagging, uploading, moving, editing, or deleting documents. This fixes a pagination bug where scrolling would skip pages after performing these operations.
- Format MCP tool names from "mcp_my_server_search_docs" to
"My Server Search Docs" in the agent stream UI
- Apply same formatting in sanitizeForDisplay for inline text
- Update tool-icons.ts getToolDisplayName with MCP support
- Use service.Name instead of service.ID (UUID) for stable tool names
across MCP server reconnections
- Add (tenant_id, name) unique index on MCPService to prevent name
collisions at the DB level
- Add collision detection logging in RegisterMCPTools for defense-in-depth
- Improve truncation logic to prioritize keeping the tool name intact
Security: first-wins policy in ToolRegistry + unique name constraint
prevents tool hijacking (GHSA-67q9-58vj-32qx).
When an MCP server re-registers after long idle, a new UUID is generated,
causing tool names to change and agent tool calls to fail. Switch from
t.service.ID to t.service.Name so tool names remain stable across
re-registrations.