Type the Theme, Not the Year: How to Ask AI About Thai Stocks So the Answer Comes From the Filings, Not From Your Question

This is a sentence from the first-quarter FY2026 MD&A of Amarin Corporations, page 2 of the filing at the SEC (market.sec.or.th — AMARIN):

"rising oil prices also led to higher electricity rates, which were significantly mitigated by our investments in solar rooftop systems"

It is a useful sentence. It tells you a company felt the energy shock, and that it had already done something about it. It came back when I typed four words — rising energy and electricity costs — with the years set in a filter.

It did not come back when I typed the question the way most people type it: "which companies were affected by energy costs in 2568?". That version returned twelve passages, and only three of them were from FY2025. The rest were from 2022, 2023 and 2024 — because the year in the sentence was treated as words to match, not as a date to filter by.

That gap is the whole subject of this article. Boom Leverage AI, the chat that went live for every signed-in reader on 11 October 2026, answers from what Thai listed companies actually wrote in their Form 56-1, MD&A and Opportunity Day sessions — and how you phrase the question decides which pages it reads. Phrase it like a search engine query and you get an answer to a slightly different question than the one you meant.

The rule in one line: put the theme in the box. Put the year, the companies and the axes in the filters. Never put the answer you expect anywhere.

Why the box asks for a theme

Open the chat and the placeholder under the input says, verbatim:

"e.g. hit by energy costs · flooding · a stronger baht"

and the hint below it:

"Write a “theme”, not a year or a direction — pick years from the filter below; the answer is what companies wrote about that theme"

That wording was a product decision, not decoration. The first version of the example questions included a year and a direction ("who warned about energy costs in 2568"), and it taught readers that the system would interpret both. It does not. Here is what actually happens to a question, in order:

1. The theme is searched by meaning across every filing in your scope. It is the same engine described in why keyword search always lost retail the trade — no executive writes "the company is in trouble", so it matches meaning, not words.
2. The years, companies and axes come from the filters, not from the sentence. A year typed in the question is just more text to match.
3. Up to 30 cited passages go to the AI, each with its ticker, fiscal period and page.
4. The AI writes an answer in which every factual line must cite one of those passages — and a validator cuts the lines that do not. That mechanism has its own write-up: how every line of an AI answer is checked against its source.

The AI never sees anything the search did not return. So the quality of the answer is capped by the quality of step 1 — and step 1 is entirely your phrasing.

The controls, and what each one costs

Before the worked examples, the levers. Every one of them changes both what the AI reads and what you pay.

A free account gets 100 search credits a day and reads the most recent filings (about five quarters back). The chat itself needs you to be signed in; plain search still works without signing in. AI usage is metered separately as a weekly token allowance that resets every Monday (UTC) — the free plan's allowance is smaller than the paid plans', and the account menu shows it as "AI usage this week".

The line the chat prints under the input says it plainly: "Every question costs credits by scope · changing scope (tickers / years / axes) re-searches the filings".

Cost discipline from model validation: a follow-up that changes nothing in the filters reads the same evidence again. If the first answer was thin, a second question on the same scope will mostly be thin in different words. Change the scope or the theme — that is where new evidence comes from.

Worked rewrite 1 — energy costs

What people type: "Which companies were hurt by energy costs in 2568?"

Why it underperforms: "2568" is not a filter. "Hurt" is a direction. In my test with no year filter, the twelve passages that came back ran from FY2022 to FY2025, and only three were from the year asked about.

The rewrite: theme rising energy and electricity costs · Years FY2025–FY2026 · Whole market · axes financials + risk (1 credit).

What it returned — every row below is verbatim, verified against the filing, with the page the quote sits on:

• NFC, Q2 FY2025, page 4 — "Utility costs increased by 5.33 MB., mainly due to an increase in electricity costs of 5.75 MB. and a decrease of 0.40 MB. in other costs." (SEC filing)
• PORT, Q1 FY2026, page 1 — "Energy prices showed an upward trend, particularly fuel and electricity prices, which represent significant cost components for the transportation and logistics sector." (SEC filing)
• KCG, Q1 FY2026, page 3 — "Soaring energy prices and elevated logistics costs have intensified the risks of a broader cost escalation." (SEC filing)
• AMARIN, Q1 FY2026, page 2 — the solar-rooftop sentence at the top of this article.

Look at what that set already contains without any AI on top: one company with a number (NFC), one with a sector-wide framing (PORT), one with a forward-looking risk (KCG), and one that says it mitigated the hit (AMARIN). Those are four different kinds of answer to "who was affected", and the AI summary can only keep them apart because the search did not pre-decide which kind you wanted.

What changes if you type the direction instead. I ran the mirror image — electricity and energy costs fell — on the same years. The results moved to a different population entirely: power producers describing their revenue. ROJNA, Q1 FY2026: "Cost of electricity sold has decreased proportionately to revenue." (SEC filing, page 2). SPCG, FY2025: "The decrease was mainly due to lower the additional electricity price subsidy (Adder), the average Ft rate and a reduction in the volume of electricity generated and sold (Energy output), as shown in the table 2." (SEC filing, page 2).

Both are true, both are relevant to "energy costs" — and neither is about a company that pays for electricity. One word of direction swapped the sector you were reading.

Worked rewrite 2 — flooding

What people type: "Did the floods last year hurt retailers?"

Why it underperforms: "last year" is relative to nothing the engine knows, and "retailers" is a sector you have decided on before seeing who actually wrote about flooding.

The rewrite: theme flooding · Years FY2024–FY2026 · Whole market · financials + risk.

• INSURE, FY2024, page 2 — "primarily due to the recognition of 97 million baht in flood-related claims (net of reinsurance) from the late-year flood event, as well as increased underwriting expenses and overall acquisition costs in line with the growth in insurance premiums received for the year." (SEC filing)
• HMPRO, FY2024, page 1 — "Furthermore, an earlier-than-normal rainy season resulted in heavy rainfall and flooding in some regions, impacting customer access to stores; the northern, northeastern, and southern regions were all severely affected." (SEC filing)
• CRC, Q3 FY2024, page 2 — "flooding in the northern and northeastern regions caused inconveniences for customers visiting stores" (SEC filing)
• RT, Q1 FY2026, page 4 — "Highway No. 42, Khlong Ngae Project, which was affected by flooding during late 2024 and again in November 2025." (SEC filing)

Yes, retailers wrote about it (HMPRO, CRC). But the only passage in this set with a baht figure is from an insurer, and the passage that tells you the problem came back a second time is from a construction contractor. If the question had been "retailers", you would have read two of the four and missed the two that carry the most information.

That is the general point: let the cross-section tell you which sectors to look at. Narrow afterwards — with "+ ticker" or "My list" — not before. The same idea, applied to sector symptoms, is in type the symptom and see who raises a hand.

Worked rewrite 3 — US import tariffs

What people type: "Will CPH be hurt by Trump's tariffs?"

Why it underperforms: it asks for a forecast. Filings record what management said, not what will happen; an AI asked to predict from them will either refuse or, worse, extrapolate. And it starts from one company, so you have no baseline for how unusual that company's language is.

The rewrite: theme US import tariffs · Years FY2025–FY2026 · Whole market first · then the same theme with "+ ticker".

The whole-market pass:

• CPH, Q1 FY2025, page 2 — "The impact of President Donald Trump's trade tax policies, which imposed tariffs on trading partners with high trade surpluses with the United States, including Thailand, which faced an additional 37% import tax. The result of negotiations between the Thai and US governments has not concluded yet." (SEC filing)
• PSP, Q1 FY2025, page 1 — "The main reasons include escalating trade tensions due to the U.S. government's import tax policies, which have created significant barriers to international trade and global transportation, particularly shipping from China to the U.S." (SEC filing)
• UREKA, Q3 FY2025, page 3 — "Contributing external factors include trade policy uncertainty, U.S. import tariff hikes, decelerating economic growth in China, and intensifying competition in the plastics market" (SEC filing)
• EVER, FY2025, page 1 — "the current economic uncertainty, as well as the U.S. import tax increase measures, high household debt, and banks being cautious in approving loans, including high interest rates" (SEC filing)

Now the CPH sentence has context. Several companies name the tariff; CPH is the one in this set that writes down a rate and says the negotiation "has not concluded yet". That is a fact about what CPH disclosed in that quarter. It is not a prediction about CPH, and the honest follow-up is not "will it be hurt" but "what did CPH say about this in later periods" — same theme, "+ ticker", wider years. Reading one company across periods is its own discipline: tracing a management narrative across years.

Mind the adjacent result: the same tariff search also returned a line about the global minimum tax — a different tax entirely, close in meaning to "tax policy". Semantic search finds what is near your theme, and near is not always the same. This is why you read the cited passage, not just the summary of it.

The workflow, step by step

This is the sequence I use. Copy it.

Step 1 — Write the theme in three to six words. A noun phrase, in Thai or English: rising energy costs, flooding, US import tariffs, customer concentration, ค่าเงินบาทแข็ง. No year, no company, no "good/bad", no question mark needed.

Step 2 — Set the years in the filter. If you want "this year", pick the fiscal years explicitly. Your plan sets the maximum depth; the free plan reads roughly the last five quarters.

Step 3 — Start wide. Whole market first. The cross-section is the point — it tells you whether a company's language is unusual or just the weather.

Step 4 — Keep the default axes unless you have a reason. Financials + risk + auditor is the default. Add Opp Day when you want what management said in the room rather than what it wrote. Up to five axes stays at 1 credit.

Step 5 — Read the sources before the summary. While the answer is being written the chat shows only "N sources found · writing the answer…"; the sources appear after the answer finishes. Open them. Every factual line ends in a [n] that points at one.

Step 6 — Verify the line you will act on. "SEC original" is free and opens the filing on the SEC's site. "Open PDF page with highlight" (2 credits) boxes the exact quote on the page. For an Opp Day source, "Open the clip at that minute" plays the moment it was said.

Step 7 — Narrow with filters, not with words. "+ ticker" or "My list" for the names you hold; that re-searches the filings for that scope. A saved watchlist makes this one click.

Step 8 — Ask the follow-up on the same scope only when you want the same evidence read differently. "Which of these gave a baht figure?" is a good follow-up. "What about 2023?" is not — change the year filter instead.

Step 9 — Keep the result. On a DELTA plan or above, "Export Excel" downloads the cited rows. The AI summary on the scan page has "Copy Markdown" and "Download .md", which carry the summary, the scope and every source with ticker, period, page and quote. For how that summary groups companies into who gains and who loses, see the AI summary on /scan.

Three things the chat will tell you, and what they mean

"No passage in the filings matches this question — try other words or a wider year range". This is information, not failure. If nobody in the market wrote about your theme in those years, that absence is itself a finding — or your theme is phrased in a way companies never use. Try the plainer word a CFO would write.

"The answer reached its length limit — the unfinished last line was removed; type “continue” to go on from here". Long cross-sections hit the ceiling. Type "continue". The cut line is removed rather than shown half-finished, so nothing you see ends mid-sentence.

"This chat is near its length limit — consider starting a new one from here". A thread has an input ceiling. "Start a new chat from here" carries the topic over. Chat history is kept for 90 days, and you can delete a thread yourself.

There is also the line that matters most: "N lines that could not be tied to a source were cut — what is shown passed the citation check". When you see it, the system removed claims it could not trace. When no cited line survives, the whole answer is withheld and the sources are shown instead — with no quota charged.

Traps — what makes a confident answer wrong

I spent years validating models inside banks. The failures that hurt were never the ones that looked broken. They were the ones that looked fine. These are the ones I see in filings questions.

1. The direction in your question picks the evidence. "Who benefited from…" retrieves benefit language; "who was hurt by…" retrieves damage language. The energy example above shows the population can change completely. Ask for the theme and let the summary sort companies into sides.

2. A year in the sentence is not a year filter. My "in 2568" test returned mostly other years. The answer can then mix periods while sounding like it is about one. Check the period on each [n].

3. Thai and Western years. Filings label periods as FY2026, Q1 and so on; the AI is instructed to write periods exactly as the source labels them and never to convert between ค.ศ. and พ.ศ. If you see a Buddhist-era year in an answer, check that the source actually says it.

4. Near is not the same. Semantic search returns what is close in meaning. "Tax policy" pulled in a global minimum tax line on a tariff question. The citation tells you what the source actually says; read it.

5. A passage is not a forecast. Everything here is what a company disclosed in a period. "Will it be hurt" is a question no filing answers. Rephrase to "what did they say about it, and did it change".

6. Free depth is shallow on purpose. The free plan reads the most recent filings. A theme that peaked two years ago may look absent on free and present on a paid plan. "No result" on free is "no result in the last five quarters".

7. Follow-ups cost credits. Every question is charged by scope, including follow-ups that reuse sources. Ten vague follow-ups on one scope spend ten times the credits for mostly the same evidence.

Mine, from validation: the AI is a reader, not a source. The source is the page. If a decision depends on a line, open the page — it is free on the SEC's site, and it is the only thing in this workflow that cannot be wrong about what the company wrote.

Where this is weak

Coverage is what was filed and indexed. If a company did not write about a theme, or wrote about it in a section outside the axes you picked, it will not appear. Absence of a passage is not proof of absence of exposure.

The AI can still be wrong. Every answer carries the line "AI summaries can be wrong — check the cited sources below before relying on one". The validator catches untraceable numbers and wrong footnotes; it cannot catch a correct quote summarised with the wrong emphasis. The details of what it checks — and what it cannot — are in the citation validator write-up.

Opp Day sources are AI transcripts. They are labelled "transcript" on the source chip, and the evidence is the minute in the recording, not a page.

I have not tested whether any of this predicts returns. It finds and quotes disclosures. It does not rank stocks.

Start here

If you hold a stock and want one question to begin with: open the chat, type the risk you worry about most as a theme, set the last two fiscal years, run it across the whole market, then add your ticker. Read where your company sits among the others. That is the whole method. The product walkthrough is in what Boom Leverage AI is and how it answers, and the one-company version is the Health Check with an AI summary.

Self-serve: sign in at terminal.boomleverage.com — the free plan includes 100 search credits a day and the most recent filings, no card. Or start from the scan page without signing in.

For a research desk or institution: deeper history, multiple seats, bulk export and API are scoped on the Enterprise page or at contact@boomleverage.com.

Disclaimer: This article is written for education and to explain how to read disclosures that Thai listed companies make publicly. It is not investment advice, it recommends no security, and it guarantees no outcome. · Every quoted passage is reproduced verbatim from a filing published on the SEC's website and linked to its source; companies named here are named only because their own filings contain the quoted text, and appearing in a result says nothing about the merit of the security · Answers in Boom Leverage AI are generated by AI from those passages, can be wrong, and must be checked against the cited source before use · Results, credit prices and plan limits described here were measured on 11 October 2026 and may change · No relationship described here has been backtested against market outcomes · Investing carries risk; past results do not guarantee future ones