fix(chat): improve security, concurrency, and context handling for RAG chat/LLM

[marevol]

Summary

• Add security hardening: session ownership verification, prompt injection prevention with delimiter markers
• Add semaphore-based concurrency control for LLM requests with configurable limits
• Make context size prompt-type aware for per-type defaults
• Preserve document order by search score when fetching full content
• Add new RAG configuration keys for history, highlights, and context tuning
• Refactor LlmClientManager to consolidate availability checks into getAvailableClient()
• Fix session integrity issues under concurrent access in ChatClient

Changes Made

Security

• ChatSessionManager.clearSession(sessionId, userId): New overload validates userId ownership before clearing; rejects mismatched user IDs with 404
• ChatApiManager: Session clear now passes userId for ownership verification
• AbstractLlmClient.buildIntentDetectionSystemPrompt(): Added anti-injection instruction appended to system prompt
• Evaluation prompt: User message and search query wrapped in delimiter markers (--- USER QUERY START --- etc.) to prevent injection
• generateDocumentNotFoundResponse: Document URL wrapped as [URL: ...] to prevent injection
• escapeQueryValue (renamed from escapeLuceneValue): Skip NULL characters in query value escaping

Context & Relevance

• getContextMaxChars(): Now prompt-type aware, allowing per-type default context sizes
• Preserve search score order when fetching full document content for context building
• Tighten relevance evaluator system prompt for stricter filtering

Concurrency Control

• AbstractLlmClient: Added Semaphore-based concurrency limiter initialized in init()
• chatWithConcurrencyControl() / streamChatWithConcurrencyControl(): All LLM calls now go through concurrency limiting
• Configurable via rag.llm.{provider}.max.concurrent.requests (default: 5) and rag.llm.{provider}.concurrency.wait.timeout (default: 30000ms)
• Returns LlmException(ERROR_RATE_LIMIT) when too many concurrent requests

Reliability

• ChatClient: User message added to session immediately (before LLM call) for session integrity under concurrency; trimHistory() moved to finally block
• AbstractLlmClient.addHistoryWithBudget(): Truncation fallback when newest history message exceeds budget
• buildStreamingRequest(): Simplified history budget — uses getHistoryMaxChars() directly instead of complex fixed-size calculation
• All history-adding calls now use addHistoryWithBudget() with getHistoryMaxChars()

Refactoring

• LlmClientManager.getAvailableClient(): Consolidated duplicate !available() checks across all 10+ delegation methods
• Javadoc updated from "Lucene query" to "Fess query" in IntentDetectionResult, ChatClient

Configuration

Added to fess_config.properties:

rag.chat.highlight.fragment.size=500
rag.chat.highlight.number.of.fragments=3
rag.chat.history.assistant.content=source_titles
rag.chat.history.assistant.max.chars=500
rag.chat.history.assistant.summary.max.chars=500
rag.chat.history.max.chars=2000

Plus config keys for intent history max messages, message max length, and per-provider concurrency settings.

Testing

• ChatClientTest: Updated to reflect new session message ordering (user message added before LLM call)
• AbstractLlmClientTest: Updated for method rename
• IntentDetectionResultTest: Updated accordingly

Breaking Changes

• ChatSessionManager.clearSession(String sessionId) (no-arg userId) still exists for admin/internal use, but the API-facing path now requires userId ownership match
• escapeLuceneValue renamed to escapeQueryValue in ChatClient (internal method)

🤖 Generated with Claude Code