CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

[Phase 6] - Earnings Call Analysis Integration

Architecture

• Two-Pipeline Design: Added a new Earnings_Agent backed by an offline ingest pipeline and a runtime inference pipeline, following the same separation pattern as the existing 10-K RAG system.
• Ingest Layer (core/earnings_tools.py): Fetches transcripts from Alpha Vantage (premium) or SEC 8-K filings (free fallback), normalizes them into Prepared Remarks and Q&A Session segments, extracts keyword frequency counts, and embeds everything into a dedicated ChromaDB collection.
• Inference Layer: Three new @tool functions — search_earnings_call (RAG search), get_earnings_sentiment_divergence (section comparison), and get_earnings_keyword_trends (cross-quarter keyword tracking).
• Graph Extension: Added Earnings_Agent to the LangGraph members list, planner capability map, supervisor dispatch edges, and summarizer prompt.

Added

• core/earnings_tools.py — Combined ingest + inference module for earnings-call data.
• scripts/ingest_earnings_calls.py — CLI tool for batch ingestion (--tickers AAPL --quarters Q1-2025 Q2-2025).
• Earnings-call example button in the Streamlit sidebar for quick testing.
• ALPHA_VANTAGE_API_KEY config setting in .env.example and core/config.py.
• Earnings Call Insights section in the Investment Memo (Summarizer) when earnings data is present.

Changed

• Planner Prompt: Extended with Earnings_Agent capability mapping for earnings-call queries.
• Summarizer Prompt: New ## Earnings Call Insights memo section for divergence and keyword findings.
• Streamlit Sidebar: Now detects and displays _earnings ChromaDB collections alongside _10k collections.

[Phase 4] - Multi-Agent LangGraph Integration & SEC Pipeline Hardening

Architecture Overhaul

• Transitioned to LangGraph: Replaced the legacy while loop with a deterministic StateGraph "Planner-Executor" architecture.
• Dual-Node Governance: Separated routing logic into a stateless Planner (generates JSON task arrays) and a stateful Supervisor (manages task queues), eliminating LLM cognitive overload and infinite routing loops.
• Separation of Concerns: Split worker capabilities into three strict React Agents: Quant_Agent, Fundamental_Agent, and Sentiment_Agent.

Added

• The "Honesty Guardrail": Implemented programmatic checks in make_worker_node to verify ToolMessage execution. If an agent attempts to answer without triggering a tool, the output is blocked to prevent hallucination.
• Strict Capability Matrix: Updated the Planner prompt to explicitly map query types (e.g., "Risks", "Supply Chain") to specific RAG tool workflows.
• Broad Query Protocol: Added fallback logic for the Planner to execute standard Quant/Sentiment tasks when users ask for "general info" on a ticker.
• Pydantic Enum Enforcement: Added strict args_schema to SEC tools using Literal types to prevent the LLM from hallucinating invalid XBRL tags.

Fixed

• The SEC "2010 Bug": The SEC API returns unordered historical data. Added Pandas-based datetime sorting, filing deduplication, and a 2-year lookback filter to ensure only modern data is served.
• The SEC "Missing Revenue" Bug: Implemented recursive fallback logic to try RevenueFromContractWithCustomerIncludingAssessedTax if the standard Revenues GAAP tag returns 404 (fixing data retrieval for MSFT, AAPL, etc.).
• ChromaDB Deprecation: Updated imports to langchain_chroma and improved SEC HTML <DOCUMENT> regex parsing for cleaner 10-K embeddings.

[Phase 5] - API Streaming, Web UI & Containerization

UI & Architecture

• FastAPI Backend: Exposed the LangGraph state machine via an asynchronous GET /chat/stream utilizing Server-Sent Events (SSE).
• Streamlit Frontend: Built a responsive agentic UI (streamlit_app.py) that visually streams the intermediate ReAct reasoning blocks into dynamically expanding dropdown menus.
• State Persistence: Rewrote the UI memory loop to preserve "Agent Thoughts" sequentially so historical messages contain dropdown logs linking exactly to how the agents generated their specific conclusions.
• Docker Migration: Shipped Dockerfile.api and Dockerfile.ui bridged via a custom network inside docker-compose.yml, successfully moving the application payload off MacOS and onto standardized, immutable infrastructure.
• LLM Engine Swap: Migrated the brain architecture permanently to Groq's llama-3.1-8b-instant and llama3-70b-versatile endpoints for incredibly fast serverless inference.

Performance & Token Thrashing Fixes

• LRU Caching: Prevented the SentenceTransformers model from re-instantiating on disk during every RAG pipeline call, silencing repetitive console logs and saving heavy JVM/Python CPU overhead.
• Groq 'Token Ghosting' Fix: Squashed a massive token quota error ("Requested 17k tokens") by explicitly declaring max_tokens=800. This prevented Groq's load balancer from assuming maximum context window limits and throttling free-tier TPM budgets.
• ReAct Infinite Loop Resolution: The Fundamental agent historically trapped itself looping search_10k_filings tools when ordered to strictly "only output data". Bridged the behavior by injecting a hard "Stop once data is fetched" logic trap inside the system prompt.
• Docker Hot-Reloading: Injected - ./:/app mapping masks and explicitly overrode the Uvicorn execution loop with --reload to support real-time Python development without tearing down the containers.

Next Steps 🚀

• Final push of the finagent backend/frontend images to Google Cloud Run.