VJRA.US / GLOBALEDGE

GlobalEdge Financial Intelligence

← Research Library
Agent: Greeter

Enter your API Key

Your Anthropic API key is used directly in the browser to call Claude. It is stored only in this browser session and never sent to any server.

Don't have a key? Get one at console.anthropic.com

01 · Business Problem
The Problem
GlobalEdge Brokerage has 400 brokers serving 180K clients across 12 countries. Each morning, 60–80 financial news articles and SEC filings are published. Brokers can only read 20–30% before their first client call — creating a structural intelligence gap. When clients challenge recommendations, brokers can't cite sources — exposing the firm to trust erosion and compliance risk.
The Solution
A no-code RAG-based market intelligence system built with n8n that ingests financial news, stock prices, and SEC filings into Pinecone. Brokers query in plain English and get sourced, grounded answers with LLM-as-a-Judge quality evaluation on every response.
02 · Solution Description
Why Not Manual Research?
70–80% of morning content goes unread
Brokers can't cite sources when challenged
Compliance review happens after client calls, not before
Why RAG + LLM-as-Judge?
Every answer grounded in retrieved context
Source citations (article, filing, date) on every response
Automated quality scoring — Relevance, Groundedness, Accuracy, Completeness, Risk Awareness
03 · By the Numbers
15
LLM-Judge Evaluations
3
Retrieval Configs Tested
5
Financial Test Cases
8.96
Best Avg Score /10 (Config A)
1,001
Chunks Generated
3/5
Test Cases Won by Config A
04 · Technical Architecture
Orchestration
n8n
No-code workflow automation
Vector DB
Pinecone
Index: global-edge
NS: globaledge-rag
Embeddings
text-embedding-3-large
1024 dimensions
OpenAI
Chunking
Recursive Text Splitter
500 chars, 50 overlap
LLM
GPT-4.1, temp 0
3 configs A/B/C tested
(top-k 5/8/10)
Evaluation
LLM-as-a-Judge
GPT-4.1, 5 metrics
15 runs → n8n DataTable
Ingestion Run
3 docs → 1,001 chunks
→ 2,444 embeddings
05 · System Architecture n8n workflows: ingest → QA RAG → evaluate
Workflow 1 — Data Ingestion
📰
Financial News CSV
global_news.csv
📈
Stock Prices CSV
stock_price_details.csv
📄
SEC 10-Q PDF
sec_filings_10q.pdf
Document Loader Text Splitter (500/50) Embeddings Pinecone
Workflow 2 — QA RAG + Evaluation
Chat Trigger QA Chain (GPT-4.1 + Pinecone, 3 configs A/B/C) Merge + Pinecone Search LLM-as-Judge Extract Metrics DataTable
Source Manifest — 3 Data Files
📰
Financial Newsglobal_news.csv
500 articles — Fed policy, oil, crypto, tech earnings
📈
Stock Pricesstock_price_details.csv
Daily OHLCV — AAPL, TSLA, GOOGL, BTC-USD, Gold
📄
SEC Filingssec_filings_10q.pdf
126-page 10-Q filing — Alphabet, Apple, Tesla
n8n Workflow Nodes
Ingestion Workflow
CSV/PDF Loader → Recursive Text Splitter (500/50) → OpenAI Embeddings → Pinecone Upsert
QA Workflow
Chat Trigger → QA Chain (GPT-4.1, Config A/B/C) → Pinecone Retriever (top-k 5/8/10) → Merge → LLM-as-Judge → Extract Metrics → DataTable
LLM-as-a-Judge Metrics
Metric What It Measures
RelevanceDoes the answer address the user's question?
GroundednessIs every claim supported by retrieved context?
AccuracyAre facts, numbers, and dates correct?
CompletenessDoes it cover all relevant aspects?
Risk AwarenessAre financial risks and caveats mentioned?
Verdict Pass / Needs Improvement
System Prompt — QA Chain
"You are a financial intelligence assistant for GlobalEdge Brokerage. Answer questions using ONLY the provided context from financial news, stock prices, and SEC filings. Always cite your sources (article title, filing section, or data date). If the context doesn't contain the answer, say so clearly. Highlight any financial risks or regulatory concerns."
Key Design Decisions
Why n8n?
No-code platform accessible to non-engineers. Visual workflow debugging. Built-in Pinecone + OpenAI integrations.
Why LLM-as-Judge?
Automated quality assurance on every response. No human review bottleneck. 5-metric scoring catches hallucination, gaps, and missing risk flags.
Why top-k=8 (Config A)?
Not assumed — measured. A controlled A/B/C study (top-k 5/8/10, 15 LLM-judge runs) found top-k=8 won 3 of 5 test cases; going deeper (top-k=10) added noise and lowered groundedness.
Why text-embedding-3-large?
Best-in-class retrieval accuracy for financial text. 1024-dim balances quality vs storage cost.
🧑
You
G
Greeter
N
Market
News
S
Securities
E
Executive
Response
Ask anything — the right specialist responds automatically
How Agents Transfer
Greeter
Routes to Market News, Securities, or Executive based on your question
News
Transfers to Securities for stock/filing details. Transfers to Executive for synthesis.
Securities
Transfers to Market News for market context. Transfers to Executive for synthesis.
Executive
Transfers to Market News or Securities if you drill into specifics.
Agents & Roles
GreeterWelcomes you, detects intent, routes to the right specialist
Market NewsFinancial news — Fed policy, oil prices, crypto, tech earnings, market sentiment
SecuritiesStock prices, SEC filings, 10-Q data, revenue breakdowns, risk factors
ExecutiveSynthesized morning briefing, client recommendations, portfolio risks — full KB access
How Responses Are Evaluated
Answer LLM-as-Judge (GPT-4.1) 5 Metrics DataTable
Every response from the QA RAG chain is automatically evaluated by a second GPT-4.1 call acting as an LLM-as-a-Judge, scoring Relevance, Groundedness, Accuracy, Completeness, and Risk Awareness (1–10 each) and rendering a Pass / Needs Improvement verdict. Results below are the actual graded results from 15 real executions (5 financial test cases × 3 retrieval configurations), stored in an n8n DataTable.
Experimental Design & Rationale
Objective: Evaluate the impact of retrieval depth (Top-K) and response length (Max Tokens) on answer quality while keeping all other parameters constant.
Evaluation Method: Each configuration was evaluated across five financial test cases using an LLM-as-a-Judge framework that scored responses for Relevance, Groundedness, Accuracy, Completeness, and Risk Awareness.
Experimental Variables: Only Top-K Retrieval and Max Tokens were varied across Configurations A–C. All remaining RAG parameters were held constant.
Selection Rationale — these two parameters directly influence the two primary stages of a RAG pipeline:
  • Top-K Retrieval determines how much relevant context is retrieved from the vector database.
  • Max Tokens determines how much of that context the LLM can incorporate into its generated response.
  • Varying only these parameters enabled a controlled evaluation of how retrieval depth and response length affect answer quality while minimizing the influence of other factors.
Total Evaluations: 15 (5 Test Cases × 3 Configurations)
Parameter Config A Config B Config C
Chunk Size500500500
Chunk Overlap505050
Top-K8510
Temperature000
Top-p111
Max Tokens10008001500
Test Case Config A Config B Config C
Test Case 1 — Amazon
Test Case 2 — Apple
Test Case 3 — Bitcoin
Test Case 4 — JPM
Test Case 5 — Nvidia
Results Summary — 15 Runs (5 Test Cases × 3 Configs)
Config Top-K Max Tokens Avg Overall Score Test Cases Won
A — Recommended 8 1000 8.96 / 10 3 of 5
B 5 800 8.76 / 10 1 of 5
C 10 1500 8.64 / 10 1 of 5
All configs: chunk size 500 / overlap 50, temperature 0, top-p 1 — only Top-K and Max Tokens were varied.
📋 Click a Test Case to Compare Config A / B / C
TC1: Based on this week's news sentiment around Amazon, would you classify the current signal as bullish or bearish, and does the price data support that?
TC2: Should I invest in Apple right now? I heard Tim Cook is stepping down — is that a red flag? How's the stock been doing?
TC3: Is now a good time to buy Bitcoin? News is saying it just crossed $78k — am I too late?
TC4: J.P. Morgan is calling for the S&P 500 to hit 7,600 on AI earnings. Is that realistic? Should I buy in now or wait?
TC5: Nvidia is leading the AI rally. How has NVDA's stock been performing compared to the news sentiment? (evaluated after a QA system-prompt refinement)
Key Experimental Findings
Configuration A (top-k=8, 1,000 max tokens) delivered the strongest overall performance in Test Cases 1–3 and the highest average score (8.96/10) across all 15 runs. Configuration C achieved the highest score only in Test Case 4, and its extra retrieval depth (top-k=10) hurt Test Case 5 enough to trigger a Needs Improvement verdict. Configuration B produced the best balance of quality and efficiency for Test Case 5. Across all five test cases, increasing retrieval depth (Top-K) and response budget (Max Tokens) alone did not reliably improve answer quality — in several cases it slightly reduced groundedness and accuracy. QA system-prompt refinement (applied for Test Case 5) had a larger effect on evaluation quality than retrieval-depth tuning did.
VP_GlobalEdge Data Ingestion — n8n Workflow JSON
Manual trigger · 3 file loaders · Pinecone upsert
How to Import into n8n
01 · Download
Pick a workflow above, then click "Download .json".
02 · Import
In n8n: Menu → Workflows → Import from File. Select the downloaded JSON.
03 · Credentials
Connect your own OpenAI and Pinecone credentials on the corresponding nodes.
04 · Run Order
Run Data Ingestion first to populate Pinecone, then run any of the 3 QA RAG configs against it.