perf: Downgrade model to 1.5B for faster CPU inference

style: Reverse session list order and rename to Session N

feat: Enhance citation format with text snippets

perf: Optimize Latency (Remove query translation, async cache, tune LLM threads, limit tokens)

feat: Phase 4 Upgrade - RAG Architecture Stability (mMARCO, pdfplumber, true recursive chunk, 3B LLM)

feat: Enhance RAG LLM prompt with few-shot examples to enforce polite and complete sentences

fix: Resolve AttributeError by correctly fetching session chunks from Global VectorStore in Phase 2 architecture

feat: Implement Phase 3 - Optimize inference engine with llama.cpp

feat: Implement Phase 2 - pyvi tokenizer, Global FAISS Index, and Hierarchical Summary

feat: Implement Phase 1 - Semantic Cache, Observability Logging, and Citation Verification

Revert hardcoded samples and set intent threshold to 0.70

Fix router false-positive chitchat classification by increasing threshold

Optimize LLM lazy loading and refine RAG prompt for full sentences

fix: increase max upload file size limit from 15MB to 100MB for large annual reports

feat: production-hardening, upload validation, admin auth, threadpool, logging & HF dataset repo integration

Implement OUT_OF_SCOPE intent separation, multi-signal relevance check, and strict Grounding System Prompt

Implement Negative Rejection and Grounding Guardrails to suppress fake citations for out-of-document queries

Implement UNKNOWN state Decision Engine with continuous document relevance score and zero hardcoded rules

Prioritize CHITCHAT intent for short greeting queries in Decision Engine

Flatten vector arrays to 1D in _cosine_similarity calculation

Fix Embedder method name from embed_query to encode_query

Implement Production Hybrid Decision Engine with Semantic Vector Centroid Router and Relevance Signal

Fix VectorStore method name from similarity_search to search_session in router pre-check

Expand LLM context window max length to 6500 tokens (~5000 words)

Add anti-bias protection to ensure queries without explicit summary intent never misclassify into META_DOC

Implement Vector Similarity Pre-Filter and Uniform Page Sampling for 100+ page documents

Expand router context window by sampling 6 chunks across full document

Implement Dynamic Document Context-Aware LLM Intent Router

Finalize smooth LLM reader responses and pure Agentic LLM intent router

Remove all hardcoded rules in router.py and sync physical disk session_docs in rag_pipeline.py

Fix router logic so document-backed sessions route queries to RETRIEVAL

Remove AI avatar icon box from bot message bubbles and indicators

Implement Two-Level Map-Reduce Hierarchical LLM Summarization for META_DOC intent

Use pure natural context understanding for LLM Router without hardcoded prompt constraints

Refine LLM router prompt rules for sharp RETRIEVAL intent classification

Merge OUT_OF_SCOPE into CHITCHAT for streamlined 3-intent architecture

Refactor router.py to pure Agentic LLM Semantic Intent Classifier

Prioritize RETRIEVAL intent for specific data and fact queries

Make chitchat LLM response dynamic with system prompt and temperature=0.7

Fix GGUF native loading dependency, fix mobile text overflow, and remove footer text

Use native Transformers GGUF loader to fix C++ compile delay in Docker build

Upgrade Local LLM to GGUF (llama.cpp) format for 5x-7x faster CPU inference

Add accelerate dependency and fix Local LLM loading error

Fix intent classification and emotional response for chitchat queries

DocBrain RAG official release for HF Space