Skip to content

LLM_EXAMPLES.md

Latest commit

 

History

History
92 lines (72 loc) · 3.89 KB

LLM_EXAMPLES.md

File metadata and controls

92 lines (72 loc) · 3.89 KB
Raw

Worked Examples — Real Generative-AI Runs

These are real outputs from gemini-2.5-flash, committed to the repo. Each report's first line is a runtime provenance banner produced in shared/llm/client.py (banner()), not a hardcoded string — the varying latencies are the actual API round-trips.

Reproduce:

export GEMINI_API_KEY=...      # free tier: https://aistudio.google.com/apikey
python demo.py

Case 1 — Earnings Call Intelligence Tracker

Input — case_1_earnings_tracker/data/itub4_q1_2026.txt (Itaú Unibanco, an Ibovespa constituent), plus data/analyst_questions.txt and the prior quarter data/itub4_q4_2025.txt for the quarter-over-quarter comparison.

Provenance banner (real run):

[LLM] gemini:gemini-2.5-flash (1 attempt(s), 9719 ms)

Selected output (full file: case_1_earnings_tracker/outputs/):

  • Management tone: cautiously optimistic (confidence 0.80).
  • Surprise score: 6/10, justified by management's explicit admission that "the conditions from the past to now are worse than the beginning of the year".
  • Top-3 analyst questions with graded answer quality High / Medium / High (the model differentiates — it does not blanket everything one grade).
  • Guidance changes vs. prior quarter (4 detected), e.g. "Reaffirmation of annual expense growth guidance at 3.5% (midpoint)" — proves the temporal comparison actually used the Q4 2025 transcript.

Grounding proof. The model is instructed to quote verbatim. The first evidence string in analysis.json is:

"The central point this quarter is that I will place somewhat greater emphasis on the credit quality of our portfolio…"

This is an exact substring of the source transcript (verified programmatically: evidence in transcript == True). The red-flag quotes are held to the same verbatim standard.

Case 2 — Macro Scenario Engine

Input — case_2_macro_engine/data/scenario.txt:

The Central Bank unexpectedly raised interest rates by 2 percentage points. Inflation remains persistent and economic growth expectations have been revised downward. Credit conditions are becoming tighter and consumer spending is slowing.

Provenance banner (real run):

[LLM] gemini:gemini-2.5-flash (1 attempt(s), 7647 ms)

Output (499 words, within the 500-word limit; full file: case_2_macro_engine/outputs/report.md):

  • Top benefited sectors — Financials (Large Banks), Utilities (Electricity), Basic Materials (Exporters), Healthcare (Defensive), Telecom — each with a transmission mechanism, e.g. "Higher benchmark interest rates directly increase net interest income for large banks…"
  • Top hurt sectors — Retail (Discretionary), Construction & Real Estate, Technology (Growth), Consumer Staples (Leveraged), Airlines & Tourism.
  • Tickers (real B3 names, justified by company traits):
    • Positive: ITUB4 (rate-sensitive lender), VALE3 (USD-revenue exporter), EGIE3 (regulated, inflation-indexed utility).
    • Negative: MGLU3 (discretionary retail), CYRE3 (mortgage-exposed builder), CVCB3 (discretionary travel).
  • Top-3 risks to the thesis — e.g. inflation proving more persistent and forcing deeper hikes; a sharp commodity-price decline; fiscal/political risk premium.
  • Confidence: 8/10, with a rationale tying the score to how direct the transmission channels are.

This output demonstrates the depth a rules engine cannot match: the model surfaces non-obvious, B3-specific names (EGIE3, CVCB3) and explains the channel, not just a label.

Why this matters

A reviewer can open either outputs/report.md, see the [LLM] banner, and read analysis that is grounded (verbatim quotes), calibrated (justified scores), and nuance-aware (graded answer quality, transmission mechanisms) — the exact qualities the use case asks Equity Strategy tooling to deliver.