Ask the Whole Thai Market a Question and Get an Answer You Can Audit — Boom Leverage AI, Where Every Line Ends in a Source You Can Open

Here is one sentence that PTT wrote about its own 2025, in its management discussion, on page 2:

"Thai Baht appreciated against US$, with the average foreign exchange rate strengthening from 35.5 Baht/US$ in 2024 to 33.0 Baht/US$ in 2025, resulting in lower revenues denominated in US$." — PTT, FY2025, MD&A, page 2 · original filing at the SEC

That is not news. Everybody knew the baht was strong in 2025. What almost nobody can answer quickly is the follow-up question that actually matters to a portfolio: which other listed companies said the same thing about themselves, in their own filings, and which of them said the opposite?

As of today you can type that question in Thai or English and get a written answer back — and every line of that answer ends in a number like [3] that opens the page it came from. This article explains what that tool is, how it decides what you are allowed to see, and why I built it the way a bank's model-validation team would demand rather than the way a chatbot is usually built.

What launched, in one paragraph

Boom Leverage AI is a chat inside Boom Leverage Terminal. It went live for every signed-in account on 11 October 2026. You ask a question; the Terminal searches the filings first — Form 56-1 One Report, MD&A, auditor reports, and our AI transcripts of SET Opportunity Day sessions — and only then lets a language model write, using nothing but the passages that search returned. Every sentence that states a fact must carry a numbered citation. A checker reads the draft before you do and removes any line it cannot tie to a source. You see what survived, a note saying how many lines were cut, and the sources underneath.

The one-sentence version: it is not an AI that knows things about Thai stocks. It is a search engine over what Thai companies wrote about themselves, with a writer bolted on top that is not allowed to say anything the search did not hand it.

What it is not — on purpose

I want to be blunt about this, because the word "AI" next to the word "stocks" makes people expect the wrong product.

If you want an opinion on whether a stock is cheap, this is the wrong tool and I would rather tell you that here than have you discover it on your third question.

Why not just drag the 56-1 into a chatbot?

Because that is the exact behaviour the Bank of Thailand's AI guidelines tell financial institutions they cannot rely on, and I wrote about why in You threw a Form 56-1 into ChatGPT and trusted the summary. The short version: a general chatbot will happily give you a number, and when you ask where it came from, the honest answer is often "nowhere". I measured how that goes wrong — units, periods, the wrong subsidiary — in AI fabricates financial numbers better than you think.

There is also a scale problem that no upload solves. One 56-1 is 120-plus pages. The question in the hook is not about one company — it is about the whole market. You cannot upload nine hundred filings into a chat window, and if you could, you would have no way of knowing which page each claim came from.

So the design question was never "how do I make the AI smarter". It was "how do I make it impossible for the AI to say something that is not on a page I can show you?"

How one question travels — five steps

This is the actual order the Terminal executes, not a marketing diagram.

1. You set the scope before you ask. Three choices sit under the box:

• Where: the whole market, SET50, or your own list of tickers.
• When: a range of fiscal years. You pick years here, not in the question — the box asks you to type a theme ("hit by the stronger baht", "energy costs", "flooding"), not "who was hit in 2568". I explain why in Ask the theme, not the year.
• Which evidence: the analysis axes — financials, risk, auditor, Opp Day and the others your pack opens.

2. The Terminal searches the filings. This is the same search engine the rest of the Terminal runs, with the same citation gate: a passage is only eligible if its quote was matched back to the source document.

3. Up to 30 sources go to the writer, numbered. The model receives a numbered list — ticker, fiscal year, period, axis, page — plus the quoted text and some surrounding context from the same document. That numbered list is the only thing it may cite.

4. The model writes, citing [n]. The rule it is given is simple: every sentence that carries a fact — a number, a company, a claim — ends with at least one [n] that points at a source on the list.

5. A checker reads the draft before you do. Lines with no citation, citations to a number that is not on the list, or a figure that does not appear in the source the line cites are flagged. Some flags get one repair pass (a wrong footnote can be moved). A number that appears in no source is never "repaired" — that line is cut. The whole mechanism is in How the validator checks every line.

What you see at the end is: the surviving answer, a line such as "cut 2 lines that could not be traced to a source" if anything was removed, and the sources underneath, collapsed until you open them.

Model-validation view: the checker is the product, not the model. Models will change. The rule "a line that cannot be traced to a page you can open does not reach the reader" is what stays fixed — and it is the same rule a validation team applies to any model whose output somebody acts on.

What the baht question actually returns

I ran the theme "ค่าเงินบาทแข็งค่า" (the baht strengthening) against financials and risk, FY2025–FY2026, through the same search the chat uses. Below are five of the rows that came back, quoted exactly as the filings wrote them, each with its page. These are the kind of numbered sources the writer would receive.

[1] PTT · FY2025 Q3 · financials · page 2

"Thai Baht appreciated against US$, with the average foreign exchange rate strengthening from 35.9 Baht/US$ in 9M2024 to 33.3 Baht/US$ in 9M2025, resulting in lower revenues denominated in US$." Open at the SEC

[2] CM · FY2025 Q3 · risk · page 1

"appreciation of the baht, which strengthened from the average exchange rate of foreign sales of 34.87 baht per dollar to 32.17 baht per dollar, or an average appreciation of more than 7.7%" Open at the SEC

[3] MEGA · FY2025 Q3 · risk · page 4

"Normalizing for currency impact due to appreciation of Thb to USD YoY, Mega We Care TM business revenue grew at low double digits in other markets in 9M 25." Open at the SEC

[4] ALUCON · FY2026 Q2 · risk · page 1

"The U.S. dollar remained strong against major currencies, supported by the resilience of the U.S. economy and relatively high interest rates. This benefited exporters generating revenue in U.S. dollars, while importers continued to face higher costs resulting from exchange rate volatility." Open at the SEC

[5] PTT · FY2025 · financials · page 2 — the sentence in the hook above.

Notice what is already visible from five rows, before any AI writes a word: three companies describe the strong baht as a drag on reported revenue, and one ALUCON passage from a later period describes the dollar as strong — a different moment in the same currency story. A good answer has to keep those apart by period. A bad one blends them into "the baht hurt exporters", which is true-ish and useless.

What a cited answer looks like

The block below is hand-composed by me from the four quoted rows above, for illustration only. It is not a captured output from the chat, and the chat's wording on the same sources will differ. It shows the shape the rules force:

Every figure in that table appears, digit for digit, in the source the row cites. That is the test the checker applies to every line the model writes. Change "33.3" to "33.0" on the PTT row while keeping [1], and the checker would flag it — 33.0 is a real PTT number, but it is in source [5], not [1]. That is a wrong footnote, and it gets repaired or cut, never shown as-is.

Following up, and when it searches again

After the first answer you can keep asking in the same thread. Two rules decide what happens:

• Same scope = same sources. A follow-up like "which of these mention hedging?" reuses the sources already found. It still charges the search credits for that scope — every question does — but it does not go back to the filings.
• Change the scope = a fresh search. Add a ticker, move the year range, switch axes, and the Terminal searches the filings again and hands the writer a new numbered list. Citations from the earlier answers are stripped out of the conversation history at that point, so the model cannot carry an old [2] into a new list where [2] means something else. That exact bug happened once in production on launch day; the fix is the stripping.

Opp Day sources behave differently on purpose. An Opportunity Day row is speech, not a document. Its card opens the recording at the minute management said it (and links the Exchange's Opp Day page), not a PDF. The "AI transcript" label on its [n] is set from the source's own metadata — the writer is not allowed to add or remove that label. More on how those transcripts are made in The Thai earnings call nobody reads.

What it costs

Two separate meters, both visible in the account menu.

• Every question charges search credits by scope, starting at 1. Up to five analysis axes in one question costs 1 credit; more than five costs 2. A wider scope adds +1 per additional 25 companies selected.
• The AI meter is weekly and counts what the model writes. It resets every Monday (UTC).
• Opening a PDF page with the quote highlighted costs 2 credits. Opening the original at the SEC is free — that link goes to the regulator's own site.
• If the answer is withheld because nothing could be traced, or the AI fails or times out, you are shown the sources instead and the AI meter is not charged. If more than half of the lines were cut, the AI output is not charged either — a mostly-wrong answer should not cost you, and the system should not be rewarded for guessing.

The full price ladder, and why the packs differ by history depth rather than by how much of the market you can search, is in what 4–6 years of memory costs.

Your history, and deleting it

Chat history is kept for 90 days, then purged automatically. You can delete any thread yourself at any time, and the deletion is permanent — the sources and answers in it go with it. You need to be signed in to use the AI; searching the Terminal without signing in works as it always has.

Five traps — what I would tell a validation committee

I spent years on the side of the table that has to sign off on models other people want to use. These are the five things I would raise about this tool if someone else had built it.

1. A cited line is traceable, not true. The checker proves that a line's numbers and wording come from the source it cites. It does not prove the company was right, that the period is the one you care about, or that management's framing is fair. The citation lets you check; it does not do the checking for you.

2. Search decides what the writer sees. If a relevant passage did not make the top 30 for your question, the answer cannot mention it. A short answer can mean "nobody wrote about this" or "the question was phrased so the search missed it". Rephrase as a theme and widen the years before concluding absence.

3. Paraphrase is where meaning drifts. Numbers are checked strictly. Wording is checked more loosely — a Thai summary of an English filing shares almost no characters with it, and must still pass. So the checker catches a line plainly lifted from the wrong source, but a subtle shift of meaning in a correctly cited line can survive. Open the source on anything you would act on.

4. Periods mix easily. "FY2025", "9M2025" and "Q2 FY2026" can all appear in one answer. The writer is told to use each source's own label and never to convert years, and the checker flags a Thai year no cited source can account for — but it is still on you to notice when two lines compare different windows.

5. Opp Day is speech recognised by a machine. A transcript row is a pointer to a minute of a public recording, not a certified record. Listen to the minute before quoting it to anyone.

The line I hold myself to: if a reader cannot get from any sentence on the screen to the page or minute it came from in one click, that sentence should not be on the screen. Everything in the pipeline — the numbered list, the checker, the cut-line counter, the no-charge rule — exists to keep that promise.

Where it is genuinely weak today

• It is new. It went to every account on launch day after a short period open only to admins. Several rules in the checker were added from production cases in the first 48 hours. More will be.
• Long answers hit a ceiling. If an answer reaches the length limit, the unfinished last line is removed and you are told; type "continue" to get the rest.
• Long threads hit a ceiling too. A very long conversation eventually reaches its size limit; the Terminal offers to start a new thread carrying the topic forward.
• It summarises; it does not calculate. It will not compute ratios across companies or build a model for you. If the number is not in a source, the answer will not contain it.

For a single company rather than a market-wide question, the Health Check card does a related job — one company, every axis — and has its own AI summary: see Health Check with AI summary. For turning a whole result list into "who gains, who loses", see AI Summary on a scan.

The big picture

The interesting thing about Thai disclosure was never that it is hidden. It is public, it is free, and it is too large for a person to hold in their head across nine hundred companies. A language model is good at exactly that kind of reading — and bad at telling you where it read it. So the whole build is about forcing the second property onto the first: let the machine do the reading, and make every sentence it writes carry the page number.

Try it yourself: sign in at terminal.boomleverage.com — the free tier gives 100 search credits a day and 30,000 AI tokens a week, with no card. Ask the baht question above, then open every [n] and check whether I was right about what it returns.

For a research desk, bank or asset manager: multi-seat access, deeper history and bulk export are scoped at boomleverage.com/th/enterprise or directly at contact@boomleverage.com.

Disclaimer: This article is written for education and to explain a research tool. It is not investment advice, it recommends no security, and it guarantees and predicts no price or outcome. · Answers in Boom Leverage AI are written by an AI and may be wrong — check every answer against its sources before using it · Passages quoted here are reproduced from public filings at the SEC; the illustrative table is hand-composed from those passages and is not a captured product output · Opportunity Day passages are AI speech recognition of public recordings and are labelled as such · Prices, quotas and features described are as of 11 October 2026 and may change · Investing carries risk; past results do not guarantee future ones