Ask Anything & Manual Peer Selection

Two Problems, Two Solutions

Finmagine AI Advisor v2.8.0 ships two independent upgrades. They solve different problems, but both come from the same root cause: Screener.in is a data platform, not an answer platform. It shows you tables and documents. It does not answer your questions.

Problem 1 — The Research Rabbit Hole

You are looking at Innova Captab. Management mentioned their Jammu facility ramp-up in every concall for the last three quarters. You want the numbers: what is the revenue contribution from Jammu, how much capex has been invested, what is the optimal capacity and timeline? All of this is in the concall transcripts.

But Screener.in gives you a list of 12 PDF links. Not answers. You would need to open each PDF, search for relevant passages, cross-reference across quarters, reconcile different numbers. That is 45 minutes of document work before you can think about the investment question.

Problem 2 — Wrong Peers

Screener.in auto-detects peer companies using GICS sector codes. For large-cap sectors with clean sub-industry definitions — IT Services, Private Banking, FMCG — this works well. TCS gets Infosys and Wipro. HDFC Bank gets ICICI Bank and Kotak.

But for mid-cap and specialized sub-sectors, GICS granularity breaks down. Innova Captab is a pharma contract development and manufacturing organization (CDMO). It gets grouped with Divi's Laboratories and Cipla — an API manufacturer and a branded generic company respectively. Neither is a CDMO. Comparing Innova Captab's margins and capital intensity to Divi's or Cipla produces a misleading peer analysis. The right peers are Akums Drugs and Windlas Biotech.

Ask Anything — How It Works

The NotebookLM Problem Statement

When NotebookLM launched, users immediately grasped its core insight: upload documents, ask questions, get citations. It became the go-to tool for reading long PDFs efficiently. The same need exists for Indian stock research. Every investor with a Screener.in premium account has access to a company's complete document library. What was missing was a structured way to ask questions against that library from within the same research workflow.

That is the design requirement for Ask Anything. The extension already has access to all document URLs for every company through the Finmagine Data Layer. v2.8.0 bundles those URLs into a targeted Q&A prompt with strict citation-only instructions and sends it to Claude.

What the Template Does

When you open the Ask Anything template on any Screener.in company page, the extension shows:

1. Document summary pill — a count of available documents for this company: e.g., "12 concall quarters, 8 annual reports, 6 credit ratings available." This tells you before you even type a question how rich the document base is.
2. Free-form question textarea — type any question in plain English. Single focused questions work best.
3. Five click-to-fill example questions to help users who are not sure what to ask.
4. BUILD PROMPT button — generates a structured prompt with your question + all document URLs + strict FINDING format instructions.

The FINDING Format

The prompt instructs Claude to respond exclusively in a structured citation format. No narrative paragraphs, no synthesis, no generic observations — only numbered FINDING blocks. Each block has exactly two fields: Source (the document name and period) and Finding (one specific factual point with exact numbers or direct quotes).

Why Claude, Not Gemini Deep Research?

Gemini Deep Research is designed to produce comprehensive research reports. That design goal conflicts directly with the Ask Anything use case. Gemini Deep Research will receive the AMA prompt, recognize it as a research question about a company, and produce a full multi-section report regardless of format instructions. It ignores the FINDING structure because its training objective is comprehensive coverage, not targeted Q&A.

Claude is trained to follow structured output instructions. When the AMA prompt says "respond only in FINDING blocks, one fact per finding," Claude complies. The output is a clean numbered list of citations — exactly what makes the template useful for rapid document research.

For this reason, the Ask Anything template card in the extension UI shows: ✦ Best with: Claude. All other templates retain their existing BEST_WITH recommendations.

Innova Captab — Ask Anything in Practice

To demonstrate Ask Anything with a real example, we ran the template on Innova Captab (NSE: INNOCAP) with the following question:

Claude read the Q1 FY26, Q2 FY26, and Q3 FY26 concall transcripts and returned 11 FINDING blocks in under 2 minutes. Here is the complete output:

What This Demonstrates

Eleven findings, all citation-backed, all traceable to a specific quarter. No hallucinations — every number matches what management actually stated. The exact revenue progression (₹60 Cr in Q2 → ₹89 Cr in Q3), the capex figure (₹480 Cr+), the revenue potential (₹1,400 Cr+ at 65–70% utilization), and the revised guidance (₹270–280 Cr from ₹400 Cr) are all direct citations, not AI-constructed summaries.

Manual Peer Selection — How It Works

The GICS Problem

Screener.in auto-detects peers from the Peers table on each company page. The grouping algorithm uses GICS (Global Industry Classification Standard) sector codes. At the broad sector level — Information Technology, Consumer Staples, Financials — GICS is accurate. But at the sub-industry level, GICS has known blind spots for Indian mid-cap companies with specialized business models.

A pharma CDMO (Innova Captab), a pharma API manufacturer (Divi's Laboratories), and a branded generic company (Cipla) can share the same GICS sub-industry code. Their business models are completely different: CDMO revenues are contract-based with lower volatility; API revenues track commodity chemical cycles; branded generic revenues depend on prescription volume and pricing power. Comparing margins, asset turns, and capital intensity across these three as "peers" produces misleading conclusions.

Previous Behavior (v2.7.0 and Earlier)

The Peer Comparison template auto-detected all peers from Screener's Peers table, immediately initiated data fetches from the Finmagine Data Layer for each peer, and showed a status panel. There was no way to override the peer list.

New Behavior (v2.8.0)

The Peer Comparison template now has a two-step flow. Step 1 is always shown before any data fetch:

Once you click "Analyze These Peers," the template proceeds to Step 2 — the same status panel as in previous versions, showing data fetch progress for each selected peer. A small "✐ Edit peers" button in the step 2 header lets you go back and change your selection without starting over.

Use Case Examples

Why Claude for Ask Anything

The four AI platforms supported by Finmagine AI Advisor have different design goals. Understanding which platform fits which template type makes a material difference in output quality.

Gemini Deep Research's training objective is to produce comprehensive coverage of a topic, not to follow a narrowly constrained output format. Even with explicit format instructions, it interprets the AMA prompt as a research question and produces a full multi-section report. This is by design — it is the right tool for Comprehensive Analysis and Forensic Governance. It is the wrong tool for targeted Q&A with atomic citations.

Claude follows structured output instructions faithfully because instruction adherence is a core capability it is explicitly trained for. The FINDING/Source/Finding format constraint is preserved across all 11+ findings in the Innova Captab output above — Claude never drifts into narrative even for complex multi-quarter questions.

v2.8.0 in Context — The Full Template List

With v2.8.0, Finmagine AI Advisor has 13 analysis templates in total — 9 for Indian stocks on Screener.in and 4 for US/Global stocks on stockanalysis.com.

US / Global Templates (4)

Also From Finmagine