Published: August 2026 · 7 min read · Premium + BYO AI key · Both markets
1. What Is the AI Research Agent?
The AI Research Agent is a dedicated chat page — finmagine.com/research-ai/ — where you ask a plain-English question about a company and get back an answer grounded in Finmagine's own indexed evidence: concall transcripts, annual reports, credit ratings, investor presentations (India), and earnings-call transcripts (US). Every factual claim in the answer carries a citation you can click through to the real source excerpt.
A research layer, not a synthesized verdict: Analyst Debate and AI Advisor give you a structured take built from Finmagine's own computed data. The Research Agent instead retrieves the specific documents relevant to your question and answers from those — closer to "show me what the filings actually say about X" than "tell me what you think about this stock."
Premium + BYO key, both markets: Requires Premium access and your own connected AI key (OpenAI, Anthropic, Gemini, or DeepSeek) at
/account/ai-keys.php. Fully live on both India and US at the same page — switch markets with the scope selector.
2. Asking a Question
🇮🇳 India
NETWEB
Balanced (8 sources)
🇮🇳 NETWEB
What's the current order book, and what did management say about revenue guidance?
The order book stood at approximately ₹2,100 Cr as of the latest concall
[1], with an additional ₹350 Cr identified as L1 (lowest bidder, pending finalization)
[2]. Management reiterated confidence in sustaining growth momentum through FY27 during both the May and January concalls
[3].
[1] NETWEB — Concall Transcript · Q3 FY26 · confidence 92%
[2] NETWEB — Investor Presentation · Jan 2026
Illustrative — modeled on a real live-tested exchange during development, condensed for this guide
Setting the scope before you ask
| Control | What it does |
|---|
| Market | 🇮🇳 India or 🇺🇸 US — determines which corpus (India Document Archive vs US Transcript Search) is searched. |
| Symbol (optional) | Scope the question to one company. Leave blank for a broader question against the whole market corpus. |
| Source mode | Fast (5 sources) / Balanced (8, default) / Deep (12) — how many retrieved chunks feed the answer. |
The status line
After you click Ask, a status line with a spinner shows live progress (e.g. "Connecting…", then retrieval and generation steps) via a streaming connection — you're not staring at a blank screen wondering if anything is happening. The connection stays open and updates the status text as each stage completes, then renders the final answer once it's fully generated and validated.
Press Enter to send: Enter submits your question; Shift+Enter adds a new line if you're writing a longer, multi-part question.
3. Fast / Balanced / Deep — Choosing a Source Mode
| Mode | Sources (top_k) | Best for |
|---|
| Fast | 5 | A quick factual check — "what was Q3 revenue growth" — where a handful of chunks is plenty. |
| Balanced (default) | 8 | Most everyday questions — enough coverage without an oversized prompt. |
| Deep | 12 | Broad or nuanced questions spanning several quarters or topics, where maximum retrieval coverage matters more than speed or your BYO token spend. |
More sources = a bigger LLM call, on your own quota: Deep mode retrieves more chunks and builds a larger prompt, which costs more tokens on your BYO key. It's not "always use Deep" — match the mode to how much you actually need for this specific question.
4. Reading the Answer & Citations
Citation markers
Every factual block in the answer ends with one or more small superscript numbers. Click one and the page scrolls to that numbered citation card below the answer.
| Marker colour | Meaning |
|---|
| 🟡 Gold | Server-verified — the cited ID genuinely exists among the retrieved evidence for this answer. |
| 🟠 Amber | Unverified — the model referenced a citation ID that didn't check out server-side. The block is flagged as unverified rather than silently trusted; any invented citation ID is stripped before the answer is even shown to you. |
Citation cards
| Field | India | US |
|---|
| Reference | Page range (page_ref) | Speaker (speaker_ref, e.g. "Kevan Parekh, CFO") |
| Confidence | Real OCR-extraction score | Always null — clean vendor text, nothing to score |
| Source link | Links externally to the original filing (e.g. a BSE PDF) | Always links internally to the company's own Documents tab — never an external vendor URL |
Why US citations never go external: Finmagine is contractually barred from naming its US transcript vendor. Every US citation is hardcoded to an internal link and a null confidence score — this was verified against a deliberately adversarial test row before the feature shipped, and re-confirmed via DOM inspection on a real live answer.
Insufficient evidence
If nothing relevant is indexed for your question, the agent shows "No indexed evidence matched this question" plus a suggestion — and doesn't make an LLM call at all. You're not billed on your BYO key quota for a question with nothing to ground an answer in.
Uncovered aspects
If your question has multiple parts and only some are covered by the retrieved evidence, a dashed-border note lists what wasn't covered — an honest gap disclosure instead of the model quietly filling in the rest from general knowledge.
5. Threads & History
- Every conversation is a thread. The chip row above the messages shows + New plus your existing threads (flag emoji + symbol or title), with the active one highlighted.
- Click a thread chip to reopen it — full history reloads, including past answers and citations.
- A thread's market and symbol scope are locked once created. You can't append a US question into an India/HDFCBANK thread — start a new thread instead. This keeps each thread's history coherent rather than mixing markets or companies mid-conversation.
Arriving via a deep link auto-asks: if you land on /research-ai/ from one of the 4 stock-page entry points with a pre-filled question, it fires automatically — you don't need to click Ask yourself for that first question.
6. The 4 Entry Points
Rather than being just another standalone page, the Research Agent is woven directly into the stock page — 4 entry points, both markets, each deep-linking with the market and symbol already scoped:
📁
"🔬 Ask the Research Agent about SYMBOL →" — right above the coverage strip.
🕐
"🔬 Ask the Research Agent why an event happened →" — for following up on anything you spot in the event feed.
📦
"🔬 Ask the Research Agent" button in the toolbar, alongside Copy and Download PDF.
⚖️
Analyst Debate tab
"🔬 Want cited evidence instead of a synthesized verdict? Ask the Research Agent →" — the explicit contrast with Analyst Debate's synthesized take.
Same idea everywhere: whichever entry point you use, you land on /research-ai/ with the market and symbol pre-filled and ready to go — the only thing left to do is type your question.