M&A Oracle

Automated due diligence intelligence — detecting contradictions between management narrative and formal records across SEC filings, XBRL financials, earnings transcripts, patents, and litigation.

---

Demo

https://github.com/omkarnica/capstone_rag_project/raw/dev/docs/MandA%20Oracle_demo.mp4

---

The Scenario

A private equity firm is considering acquiring a tech company for $2B. Analysts need to answer: Is the CEO's public optimism consistent with what the financial filings actually say? Are there hidden liabilities in the footnotes?

M&A Oracle automates this process — routing due diligence questions to the right evidence source, cross-referencing multiple filing types, and flagging contradictions with citations.

Example query:

Identify all instances where Meta's CEO's earnings call statements about metaverse investment contradicted the CapEx guidance in their 10-K filings from 2022–2024.

The system retrieves all relevant Meta earnings transcripts, extracts CEO statements, fetches CapEx figures from XBRL data, compares narrative claims against actual numbers, and flags specific contradictions.

---

System Architecture

Request Flow

User Query
    │
    ▼
FastAPI  (/query, /due-diligence)
    │
    ▼
Semantic Cache ──► Cache Hit → Return Cached Answer
    │ Miss
    ▼
Planner  (single-hop Tier 1 / multi-hop Tier 2 / llm_direct Tier 0)
    │
    ▼
LangGraph Workflow
    │
    ├─► Router ──► sql          → NL-to-SQL → PostgreSQL XBRL
    │             ├─► filings   → RAPTOR retrieval → Pinecone
    │             ├─► transcripts → Pinecone hybrid search
    │             ├─► patents   → Pinecone hybrid search
    │             ├─► litigation → Pinecone hybrid search
    │             ├─► graph     → Cypher generation → Neo4j
    │             ├─► contradiction → multi-source sweep
    │             └─► llm_direct → Gemini (no retrieval)
    │
    ├─► Rewriter  (query reformulation)
    ├─► Retriever → Grader  (relevance filtering)
    ├─► Merge     (multi-hop sub-question fusion)
    ├─► Generator (cited answer synthesis)
    └─► Fallback  (Gemini Google Search grounding)
         │
         ▼
Audit Logger → BigQuery  +  LangSmith trace

Hybrid Retrieval

All vector routes use Reciprocal Rank Fusion over dense (Pinecone) and sparse (BM25) rankings before passing top-k chunks to the grader. Implementation: src/utils/hybrid.py.

Knowledge Graph

Neo4j graph covers Companies → Filings → Sections, Subsidiaries, BoardMembers, Patents → TechnologyDomains. src/graph_retrieval.py generates read-only Cypher via Gemini, executes it, and formats rows as evidence. src/Knowledge graph/kg.py handles ingestion.

Agentic Multi-hop Pipeline

src/nodes/planner.py decomposes complex questions into ordered sub-questions, each with a route hint. Sub-questions execute through the full LangGraph workflow and results are fused by src/nodes/merge.py before final generation.

---

Tech Stack

Layer	Technology
API	FastAPI + Pydantic
Orchestration	LangGraph
LLM	Gemini 2.5 Flash via Vertex AI
Vector store	Pinecone (ragcapstone index)
Embeddings	Pinecone llama-text-embed-v2
Reranker	Pinecone bge-reranker-v2-m3
Knowledge graph	Neo4j
Relational DB	PostgreSQL (Cloud SQL)
Cache	Redis (Cloud Memorystore)
Secrets	GCP Secret Manager
Observability	LangSmith + BigQuery audit log
Evaluation	deepeval + custom metrics
Frontend	React + Vite
Package manager	uv

---

Data Sources

Source	Content	Volume
SEC EDGAR	10-K, 10-Q, 8-K filings	Millions of documents
XBRL	Structured financial facts	Hundreds of tags per filing
Earnings transcripts	CEO/CFO Q&A, guidance statements	Thousands of calls
USPTO patents	Claims, CPC codes, assignees	13M+ patents
CourtListener	Litigation records, court cases	Federal + state courts
SEC DEF 14A	Proxy statements, board composition	Per company per year
SEC Exhibit 21	Subsidiary listings	Per filing year

---

Retrieval Strategies

Route	Trigger	Module
sql	Financial metrics, XBRL numbers, balance sheet	src/nl_sql/pipeline.py
filings	10-K/10-Q narrative, risk factors, MD&A	src/filings/raptor_retrieval.py
transcripts	Earnings call statements, CEO/CFO quotes	src/transcripts/retrieval.py
patents	IP portfolio, patent claims, CPC codes	src/patents/retrieval.py
litigation	Court cases, settlements, legal exposure	src/litigation/retrieval.py
graph	Subsidiaries, board members, ownership structure	src/graph_retrieval.py
contradiction	Cross-source due diligence sweep	src/contradictions/detector.py
llm_direct	General M&A concepts, no retrieval needed	Gemini direct

---

Evaluation Framework

Metrics

• NumericalAccuracy — fraction of expected financial figures found within 1% tolerance
• ContradictionDetection — precision/recall on flagged narrative vs. filing discrepancies
• GeminiJudge — LLM-as-judge scoring for answer faithfulness and citation quality
• RetrievalPrecision/Recall — chunk-level relevance evaluation

Ablation Study

evals/configs/ablation_configs.py defines retrieval configurations (dense-only, hybrid, with/without reranker, with/without RAPTOR). evals/runner.py runs all configs against evals/dataset/ and outputs a baseline_delta comparison to evals/results/.

Observability

All LangGraph runs are traced in LangSmith (project RAG-Capstone). Every query writes a structured AuditRecord with UUID correlation ID, tenant, user, route, latency, sources, and confidence score to BigQuery (codelab-2-485215.ma_oracle.audit_log).

---

Setup

Prerequisites

• Python 3.11+
• uv package manager
• GCP Application Default Credentials (gcloud auth application-default login)
• Pinecone API key
• Neo4j instance (URI, user, password)
• Redis instance
• PostgreSQL database ma_oracle on port 5433 (local dev)

Environment

Create a .env file at the project root:

PINECONE_API_KEY=your-pinecone-key
NEO4J_URI=bolt://your-neo4j-host:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=your-neo4j-password
REDIS_URL=redis://localhost:6379
LANGSMITH_API_KEY=your-langsmith-key   # optional, auto-fetched from Secret Manager

Secrets not in .env are resolved from GCP Secret Manager at startup.

Install

uv sync

Run

uv run uvicorn src.app:app --reload --port 8000

API docs available at http://localhost:8000/docs.

Tests

uv run pytest -q

---

Local Dev Setup

Prerequisites

• GCP access to project codelab-2-485215
• gcloud CLI installed and authenticated

Start Redis Tunnel (required before running app)

The app connects to Cloud Memorystore Redis via an SSH tunnel through the bastion host. Keep this terminal open while running the app.

gcloud compute ssh redis-bastion `
  --zone=us-central1-a `
  --project=codelab-2-485215 `
  --ssh-flag="-L 6379:10.101.67.12:6379" `
  --ssh-flag="-N"

Start App

In a new terminal:

uv run uvicorn src.app:app --reload --port 8000

---

Deliverables

Deliverable	Status
Deployed web application	✅ Live
Authentication with role-based access	✅ Tenant + user ID isolation on all requests
Multi-tenant architecture with data isolation	✅ tenant_id scoped across query, audit, and cache layers
Document lifecycle management	✅ Ingestion pipelines for SEC, patents, litigation, transcripts
React frontend — query interface + admin dashboard	✅ Vite/React app
Knowledge graph with visualization	✅ Neo4j graph, Cypher generation, entity relationship queries
RAG router with minimum 4 retrieval strategies	✅ 7 routes (sql, filings, transcripts, patents, litigation, graph, contradiction)
Agentic pipeline for multi-hop queries	✅ Planner → sub-questions → merge → generate
Multimodal handling	✅ XBRL structured data + filing text (RAPTOR) + earnings transcripts + patents
Evaluation framework (RAGAS + custom metrics)	✅ deepeval + NumericalAccuracy + ContradictionDetection + GeminiJudge
Ablation study	✅ Dense vs hybrid vs reranker configs with baseline delta scoring
LLM observability	✅ LangSmith tracing + BigQuery audit log with correlation IDs
Slack/webhook notifications	⚠️ In progress
Structured logging with correlation IDs	✅ UUID query_id on every request, BigQuery sink
GitHub repository with documentation	✅ This repository

---

Team

Name	Role
Karnica Jain	Backend, Authentication, ingestion pipelines
Ruby Gunna	RAG pipeline, LangGraph orchestration, evaluation
Nithu Arjunan	Frontend,knowledge graph,raptor tree,ingestion, deployment