Continuous enhancement: quality context panel

.env.example

@@ -0,0 +1,113 @@
+ALPHA_VANTAGE_API_KEY=
+FMP_API_KEY=
+SEC_USER_AGENT=FinGPTLocalResearch/1.0 contact@example.com
+QDRANT_URL=http://localhost:6333
+QDRANT_API_KEY=
+COLLECTION_NAME=market_docs
+OLLAMA_BASE_URL=http://localhost:11434
+PRIMARY_MODEL=qwen2.5:7b
+GEMMA4_MODEL=gemma4:e4b
+ENABLE_EXPERIMENTAL_FALLBACK=false
+EXPERIMENTAL_FALLBACK_MODEL=gemma4:e4b
+HF_TOKEN=
+HF_MODEL_REVISION=main
+
+# Localization (ko: Korean, en: English)
+OUTPUT_LANGUAGE=ko
+
+# --- Macro bundle (bonds / FX / commodities / crypto) ---
+# Register a free key at https://fred.stlouisfed.org/docs/api/api_key.html
+# Leave empty to disable the FRED half of the macro bundle; the pipeline
+# will still collect yfinance price snapshots + Google News headlines.
+FRED_API_KEY=
+MACRO_PRICE_LOOKBACK_DAYS=90
+MACRO_NEWS_QUERY_LANGUAGE=en
+
+# --- Structured data mart ---
+# Separate from data/runs.db, which remains the research-run history index.
+DATA_MART_BACKEND=sqlite
+DATA_MART_DB_PATH=data/research_mart.db
+DATA_MART_DUCKDB_PATH=data/research_mart.duckdb
+TELEGRAM_BOT_TOKEN=
+TELEGRAM_CHAT_ID=
+
+# --- Data provider policy ---
+# Default stack: direct Yahoo/yfinance + FRED + SEC + Google News RSS.
+# OpenBB is installed and checked, but news runtime is opt-in because provider
+# compatibility can vary by version. FMP is auxiliary and disabled by default.
+DATA_PROVIDER_PRIORITY=yfinance,sec,google,openbb,alpha_vantage,fmp
+ALPHA_VANTAGE_ENABLED=false
+OPENBB_ENABLED=true
+OPENBB_NEWS_ENABLED=false
+# --- Optional OpenBB Workspace custom agent adapter ---
+# Keep disabled unless you want OpenBB Workspace to discover FinGPT via
+# /agents.json and stream queries to /query.
+OPENBB_AGENT_ENABLED=false
+OPENBB_AGENT_ID=fingpt-local-research
+OPENBB_AGENT_NAME=FinGPT Local Research
+OPENBB_AGENT_PUBLIC_URL=http://127.0.0.1:8000
+OPENBB_AGENT_ALLOW_ORIGINS=https://pro.openbb.co,http://localhost:1420,http://127.0.0.1:1420
+FMP_ENABLED=false
+TRANSCRIPT_PROVIDER=fmp_optional
+
+# --- Retrieval strategy ---
+# multi_query: fan the user question into raw + risk-focused + catalyst-focused
+#              sub-queries and fuse results with RRF (default, deeper recall).
+# single:      legacy single-query semantic search.
+RETRIEVAL_STRATEGY=multi_query
+
+# --- Cross-encoder reranker ---
+# Disable to revert to pure RRF ordering. The default model is intended to run
+# locally on CPU when a compatible FastEmbed/sentence-transformers runtime is
+# available; failures are logged and the original order is preserved.
+RERANKER_ENABLED=true
+RERANKER_MODEL=Xenova/ms-marco-MiniLM-L-6-v2
+RERANKER_CANDIDATE_POOL=30
+
+# --- Fundamentals card ---
+# Single-name equities only. Timeouts/failures skip the card and keep the
+# normal RAG path running.
+FUNDAMENTALS_CARD_ENABLED=true
+FUNDAMENTALS_CARD_TIMEOUT_S=5
+
+# --- Chunked ingest ---
+# Requires re-creating the Qdrant collection when switching from legacy
+# document-level vectors. See scripts/migrate_collection.ps1.
+INGEST_CHUNKING_ENABLED=true
+INGEST_CHUNK_TOKENS=512
+INGEST_CHUNK_OVERLAP=64
+MAX_CHUNKS_PER_PARENT=1
+
+# --- Hybrid search (dense + sparse BM25) ---
+# Requires re-creating the Qdrant collection so sparse vectors are indexed.
+HYBRID_SEARCH_ENABLED=true
+SPARSE_MODEL=Qdrant/bm25
+
+# --- Topic mode / universal routing ---
+TOPIC_MODE_ENABLED=true
+ROUTER_MODEL=qwen2.5:7b
+TOPIC_RETRIEVAL_TOP_K=12
+TOPIC_MAX_RELATED_TICKERS=8
+
+# --- Optional scenario simulation layer ---
+# Disabled by default. When enabled, outputs are advisory scenario analysis
+# only, not price prediction, and are stored under execution_meta.extras.
+SCENARIO_SIMULATION_ENABLED=false
+SCENARIO_SIMULATION_MAX_PERSONAS=6
+SCENARIO_SIMULATION_MIN_PERSONAS=5
+SCENARIO_SIMULATION_DEBATE_ROUNDS=1
+SCENARIO_SIMULATION_MAX_SCENARIOS=4
+SCENARIO_SIMULATION_LLM_ENABLED=true
+SCENARIO_SIMULATION_STRICT_EVIDENCE=true
+SCENARIO_SIMULATION_FAIL_OPEN=true
+
+# --- Embedding model ---
+# Keep the default BAAI/bge-small-en-v1.5 (384-dim) to stay compatible with
+# existing market_docs collections. To opt into the larger bge-base model
+# (768-dim, better semantic recall), set BOTH variables below and then drop
+# and re-create the Qdrant collection (ingestion will auto-size on the next
+# run):
+#   EMBEDDING_MODEL=BAAI/bge-base-en-v1.5
+#   EMBEDDING_VECTOR_SIZE=768
+EMBEDDING_MODEL=BAAI/bge-small-en-v1.5
+EMBEDDING_VECTOR_SIZE=384

.gitignore

@@ -103,9 +103,12 @@ celerybeat.pid
 
 # Environments
 .env
+.env.*
+!.env.example
 .venv
 env/
 venv/
+venv311/
 ENV/
 env.bak/
 venv.bak/
@@ -139,3 +142,29 @@ FinGPT_Training_LoRA_with_Chatglm2_6b_for_beginners.ipynb
 
 # Benchmark data
 fingpt/FinGPT_Benchmark/data/*/**
+
+# FinGPT local runtime artifacts
+data/outputs/
+data/raw/
+data/runs.db
+data/research_mart.db*
+data/research_mart.duckdb*
+data/quant_lab/backtests/
+data/quant_lab/strategies/
+data/watchlist.json
+data/ai_portfolio/
+data/ai_portfolio_smoke/
+data/ai_portfolio_browser_smoke/
+data/forecast_lab/experiments/
+data/forecast_lab/data_snapshots/
+data/forecast_lab/macro_regime/
+data/forecast_lab/model_artifacts/
+data/forecast_lab/model_registry.json
+data/forecast_lab/model_registry.sqlite3*
+data/forecast_lab/jobs.sqlite3*
+data/forecast_lab/.artifact_signing_key.json
+reports/
+test-results/
+.playwright-cli/
+.ruff_cache/
+scratch/