Retail Does Not Lose to Funds on Data Speed. It Loses on Memory — And Here Is What 4–6 Years of Memory Costs (Free / ฿799 / ฿1,399)

Before the long version

• You can evaluate it before paying — a free account gets 100 credits a day with no card on file, and the database it searches is the same one every paid pack searches
• What you get: semantic search across the management discussion and analysis (MD&A) in Form 56-1 for effectively the whole market (926 companies), returning the actual text plus a link to the original filing at the SEC (Thailand's Securities and Exchange Commission) so you can check it yourself
• Where pricing starts: free, 100 credits/day, no card → DELTA ฿799/month (4 years back · 150 credits/day) → GAMMA ฿1,399/month (6 years · 350 credits/day) · annual costs the same as 10 months
• What actually differs between packs: how many years back you can look, and the daily credit quota — the database and the number of searchable companies are identical in every pack. A retail user on the free tier sweeps the entire market exactly like a fund does
• Team/institutional: seats · Excel export · API → Enterprise page or contact@boomleverage.com
• ALPHA ฿9,900 has no buy button on purpose — 10 years back · Opportunity Day · seats · export · API, scoped and priced in a conversation on the Enterprise page rather than bought from a card (spelled out plainly at the end)

Thai retail has held one belief for a long time: funds win because they hold secret data — Bloomberg screens at tens of thousands of baht a month, broker colour, research nobody else reads.

I worked model risk inside a bank for years. I sat in the rooms with those screens. The truth is more boring than that: the Form 56-1 a fund reads is the identical file you can download free from the SEC website. Same release moment. Everyone sees it at once. There is no special edition.

What they pay for is not data. It is memory.

Retail opens the latest quarter, hits the line "เลื่อนแผนลงทุนออกไปก่อน" [the capex is parked, and note they did not say for how long], and reacts on the spot — here we go, the company is in trouble — or fine, they are just being careful. One snapshot, one verdict. That is recency bias, and it is the most expensive trap in reading filings.

People who read filings for a living do not answer that question at all. They open a 4–6 year window and ask something else entirely: "How many times has this management said this exact sentence, and did they follow through the last time?"

This is a price page. I am not hiding that. But I am not selling you a feature count. I will prove it with the real thing — one company, one question — and show you that the length of your memory changes the answer. Then you decide what a longer memory is worth per month.

1. Short memory vs. long memory: same page, opposite conclusions

That last row is the entire page — and it is why the price scales on years of hindsight rather than on company count or feature list. What you are buying is a magnifying glass pointed at the past, not software.

2. How the tool works (short, before the money)

MD&A is the management discussion and analysis section of Form 56-1 (One Report) — narrative, written by management, on what happened and what is coming. It is a public document a listed company is required to disclose (the full structure of the report is in how to actually read a One Report / 56-1).

The difference here is semantic search — matching on meaning, not exact strings. You type an ordinary Thai question; the system matches it against the management discussion of effectively every listed company (926 companies), covering 2021–2026, and returns the closest passages with company, year, page number and a link to the source. Why that beats Ctrl+F is in how semantic search differs from exact-match search.

On the site you just type a Thai question into the search box; there is no code to write. The concept code below is there to show what runs underneath — and, more usefully, exactly where the packs diverge technically. The axis that sets the price is the year window you can search:

# ตัวอย่างเชิงแนวคิด — บนหน้าเว็บพิมพ์คำถามไทยในช่องเดียว โค้ดนี้แค่ให้เห็นเบื้องหลัง
from boomleverage import mdna

# แพ็กต่างกันตรง "หน้าต่างปี" ที่ค้นได้ — นี่คือแกนที่กำหนดราคา ไม่ใช่จำนวนบริษัท
hits = mdna.search("แผนขยายกำลังการผลิต", window_years=6)   # GAMMA: เห็นทั้งเส้น 2564–2569
# hits = mdna.search("แผนขยายกำลังการผลิต", window_years=4) # DELTA: 4 ปีล่าสุดของแต่ละบริษัท
# hits = mdna.search("แผนขยายกำลังการผลิต")                  # ฟรี: เฉพาะงบล่าสุด

for r in hits["results"]:
    print(r["ticker"], f"ปีงบ {r['fy']}", r["period"])   # บริษัท + ปี + ไตรมาส
    print(r["snippet_verbatim"])                          # ข้อความจริงจาก MD&A (verbatim)
    print(r["page_ref"], r["source_url"])                 # เลขหน้า + ลิงก์ไฟล์ 56-1 ต้นฉบับ ก.ล.ต.

Note that no parameter anywhere caps the number of companies — every pack fires the same question across the same full market. The only variable that moves with price is window_years. The other thing that matters is the snippet_verbatim + source_url fields: every answer carries the source text unmodified plus a link to the actual file, so you can read the full context and verify it yourself. Not an AI summary that sounds good and cannot point at where it came from.

3. The one case that explains the whole price page: KCE's new plant

I am not going to argue "more years is better" in the abstract. Here is the real thing. The question is "KCE จะสร้างโรงงานใหม่ไหม" [is KCE actually going to put the new plant in the ground] — the question every electronics-components analyst has to answer annually, and the one every holder of this stock wants answered just as badly.

This is what each level of memory looks like when you fire the same question:

Same query: "KCE จะสร้างโรงงานใหม่ไหม" [is KCE actually going to build it]

  Free  ──────────────────────────────────────────────────────────────────►  [ FY2026 Q1 ]
  (latest filing)                                                             "remains on hold"
                                                                              ⇒ Read: "it is on hold" — done

  DELTA 4 yrs  █─────────────► [ FY2023 Q3 ] ─► [ FY2024 Q1 ] ─► [ FY2025 Q3 ] ─► [ FY2026 Q1 ]
                                "postpone"       "postpone"       "postponed"      "on hold"
                                                                  + utilisation < 70%
                                "⇒ Read: deferred at least 3 times" + a trigger to watch

  GAMMA 6 yrs  ► [ FY2022 Q3 ] ─► [ FY2023 Q3 ] ─► [ FY2024 Q1 ] ─► [ FY2025 Q3 ] ─► [ FY2026 Q1 ]
                 "Bt 8,060 mn"
                 (where it starts)
                 ⇒ Read: "฿8bn project parked 4 years, ready to fire the moment demand returns"

Now the real language, word for word. Every box links back to the source.

On the free tier (latest filing only) you see exactly one sentence:

KCE · FY2026 (Q1): "With respect to the Company's capacity expansion plan, the investment in a new manufacturing facility remains on hold."

— verbatim from Form 56-1, page 2 · source filing at the SEC

On that alone, all you can conclude is that it is on hold right now — which reads like ordinary prudence. It cannot tell you how long it has been on hold, or how many times it has been put there. That is a one-quarter memory.

On DELTA (4 years) the window reaches back to FY2023 and the picture flips immediately, because none of this is new:

KCE · FY2023 (Q3): "Due to the slowdown of global market with uncertain market conditions of customers demand and the current production capacity utilization status, the company has decided to postpone the construction of new factory building for some times."

— verbatim from Form 56-1, page 5 · source filing at the SEC

KCE · FY2024 (Q1): "After considering the market situation in the electronics industry, which is still sluggish in the first half of this year, the Company has decided to postpone the construction of the new factory at the Rojana Industrial Park to around the end of the year."

— verbatim from Form 56-1, page 6 · source filing at the SEC

KCE · FY2025 (Q3): "The Company would like to inform that the plan for the construction of a new factory has been temporarily postponed. This decision reflects the Company's consideration of the current production capacity, which remains below 70%, and the strategic intent to optimize existing resources while maintaining cost efficiency and financial stability."

— verbatim from Form 56-1, page 1 · source filing at the SEC

The read changes from "on hold" to "deferred at least three separate times since FY2023" — and it arrives with a number you can work with. Capacity utilisation below 70% is the condition the company itself wrote down as the reason for not building. That is a trigger you can monitor to know when the capex comes back. This is not mind-reading management. It is the condition they published. You only need a memory long enough to see it.

On GAMMA (6 years = every year currently in the archive) you get the piece that matters most — where the whole story starts, in FY2022:

KCE · FY2022 (Q3): "The value of the investment project includes the estimated construction cost of the factory building, the installation of the facility system, the purchase of machinery, equipment and manufacturing support equipment and other expenses, total Bt 8,060 million."

— verbatim from Form 56-1, page 5 · source filing at the SEC

Now the arc is complete: a ฿8,060 million project announced in FY2022 → deferred in FY2023 → deferred again in FY2024 → still deferred in FY2025 because utilisation sits below 70% → still on hold in FY2026, with the company adding that "This readiness allows the Company to execute its expansion plan on a fast-track basis once there is a clearer and stronger demand outlook." (FY2026 Q1, page 2 · source)

Why this is money: "an ฿8 billion project parked for four years, ready to fire the moment demand turns" and "it is on hold right now" are two different recommendations off the identical document set. The gap is not skill on the part of the reader, and it is not who got the file first. It is the time window you can see — and that is the only thing the pack price buys.

4. The 2026 packs, in plain language

Read this before you decide off that table: the free column is not a trial that expires — it is a standing tier, and it searches the same 926 companies the paid packs do. What a paid pack buys is depth and quota, nothing else. Work out which question you ask most often, then buy the depth that answers it.

If you are still unsure where you sit, decide off this — it keys on the question in your head, not the size of your wallet:

Which question do you ask most often?

  "Who mentioned X this quarter?"
        └─► Free is enough  (do not pay yet, seriously)

  "Has this company said this before, and did they follow through last time?"
        └─► DELTA ฿799  (4 years = repeat behaviour + Excel export)

  "Show me the whole line, from the day they announced the project to today."
        └─► GAMMA ฿1,399  (6 years = every year in the archive)

• Excel export is a paid-tier right (DELTA and up) — the free tier searches and reads on screen without limit, but it cannot pull the full ranked result set (up to 100 rows, each with page number + SEC link) out as a file. The reason is blunt: the free tier exists so you can test until you are convinced, not to be a data pipe — and taking results into a spreadsheet to work on them is exactly what people doing this for a living pay for.
• The year window counts back from each company's own latest filing, not a fixed calendar year — companies whose fiscal year does not match the calendar (the March-close names, for instance) do not drop out of results.
• Annual = the price of 10 months (two months free) on both DELTA and GAMMA.
• Credits are charged by search scope: the whole market, or any selection of ≤25 companies = 1 credit · every additional 25 companies selected = +1 credit (50 companies = 2 credits). Which means normal use — sweeping the entire market — is the cheapest thing you can do. You are not penalised for it.
• Axes are priced in two steps, not per axis (since 11 October 2026): ticking up to 5 analysis axes (financials · risk · auditor · Opp Day · operating quality · governance · business · news) costs 1 credit; ticking more than five costs 2 credits flat. Scope and axes are added, never multiplied — 50 companies on 3 axes is 1 + 1 = 2, not 1 × 2.

If you read an earlier version of this page: there used to be a Starter pack (2 years back) as the cheapest paid tier. It is gone. The blunt reason is that not one person ever bought it — keeping a pack nobody buys on the price page only makes the choice harder, and lifting the free quota so people can actually test it is worth more. If you find a "three packs, 1/3/5 years" price table anywhere on the internet, that is the old version, and it has been replaced. Updated 29 September 2026: the free quota went up again, from 10 to 100 credits a day (the July–September launch window that handed every account DELTA for free has closed — what it was for). Depth did not change — free still stops at the latest filings and still has no export — so the reason to pay is exactly what it was: years of history, not more questions.

What the AI layer costs (added 11 October 2026)

Since 11 October every signed-in account — free included — can switch on Boom Leverage AI: a written answer over the passages a search returns, where every factual line has to end in a numbered source you can open, and lines that fail the check are cut before you see them (how the chat works · how the checker decides). It is priced on two meters, and I want both visible on this page rather than discovered at the till:

• A chat question costs the same search credits as the search it runs — by scope and axes, from 1 credit. A follow-up that stays on the same companies, years and axes reuses the sources already found; changing any of the three searches the filings again.
• An "AI summary" on top of a /scan result costs 2 credits — charged only after the checker has run, and only if no more than half of the summary's lines had to be cut. Reloading the identical summary within 24 hours costs nothing; pressing "regenerate" is a new run. (what the summary does)
• If nothing survives the check, nothing is shown and nothing is charged. The writing allowance is a fair-use ceiling, not a product — the depth of history is still what each pack actually sells.

5. Why no pack caps the number of companies (even though that would sell far better)

The easiest pricing model in financial software is selling by how many stocks you can see — free gets 10, mid gets 100, top gets the market. I know it sells better. I chose not to build it.

Not out of generosity. It would kill the only value this tool has. Semantic search is most useful precisely when you do not know in advance who in the market is talking about the thing you care about. If you have to pick 10 tickers up front, you have to guess right up front — the harder job, and the one the tool should be doing for you.

So this line holds on every pack, free included: you fire one question across the management discussion of effectively the whole market (926 companies), the same sweep a fund gets. What differs is how far back you can see. The real value of MD&A is not in a single snapshot; it is in comparison across time — who said it first, who changed posture, who said it and then did it. (A detailed cross-year read is walked through in reading MD&A across years — the PSL case)

The common follow-up: "then shouldn't I just buy GAMMA?" — no. If your working question is "who mentioned X this quarter," free or DELTA answers it more directly and far more cheaply. Choose on the question you actually ask, not on the highest price. What I recommend: start free, fire 10 real questions about your own book, and see whether the answers you want genuinely need years of history.

6. ALPHA ฿9,900 — open, and sold by conversation rather than by button

ALPHA ฿9,900/month is the desk licence: 10 years of history · Opportunity Day · multiple seats · Excel/CSV export of the whole result set · API · 1,500 credits/day. It is sold through the Enterprise page, not through a checkout page, and the access key is issued by hand once scope is agreed.

An earlier version of this article said ALPHA was a dated waitlist for Q4 2026 and would not be sold. That was true when it was written, and I am leaving the reasoning visible rather than quietly deleting it, because the reason it changed is the same standard that made it a waitlist in the first place: the two things that were missing are now measured, not promised.

• 10 years back — the searchable index now starts at FY2016, not FY2021. The raw files were always on disk (checked 2026-08-13: 94.6–98.9% of the companies that actually filed in each year); what was missing was extraction and, more importantly, proof. Findings from before FY2021 are held out of results unless a row carries an explicit verification that its quote was located in the source PDF — measured at 98.79% across the archive. A row that was never checked is withheld, not shown with a shrug.
• Opportunity Day — the ingestion pipe exists and Opp Day is now a searchable axis on every paid pack, by window: DELTA the past 4 years, GAMMA the past 6, ALPHA the past 6 (every round in the archive) — since 12 October 2026 the paid Opp Day window matches the pack's filings depth, each passage linked to the minute it was spoken. Every row is our own AI transcript of the recording, labelled as such on the card, because it is management speaking rather than a document the company wrote and filed.

What has not changed is the standard. If I had sold "10 years" while only 6 were searchable, you would have had no reason to believe the 4-year and 6-year numbers on DELTA and GAMMA either — the credibility of a price page is measured on the line that costs it money, not the line that sells. ALPHA opened when the archive could be proven, and not before.

Why there is still no ALPHA buy button: a desk licence is scoped in a conversation — seats, export volume, API rate — and a fixed monthly figure next to two self-serve cards invites a comparison that has no meaning. ฿9,900 is where the conversation starts.

7. The limits, stated plainly

Same rule I carried out of model validation: a good tool tells you what it does not do.

• It finds and quotes; it does not decide for you — semantic search returns the passages closest in meaning, and sometimes that pulls in something adjacent to your point. Your job is to click through, read the full context, and judge. A result you cannot trace back is a result you cannot use, which is why every hit carries a page number plus a link to the source Form 56-1. (why numbers from an AI go through three layers of checking)
• DELTA reaches 4 years and GAMMA 6 — the index itself runs back to FY2016, but the years before FY2021 are ALPHA depth, and even there a finding is only shown if its quote was verified against the source PDF.
• Some filings are scans read through OCR, so quoted text can carry odd seams — another reason to open the original before you act on it.
• A long memory does not make you right about the future — knowing KCE has deferred this project for 4 years tells you nothing about whether they build next year. It only stops you flinching at a single headline and starts you watching the condition the company wrote down itself.
• Not trading advice — this searches management's own words for signal and context. It does not tell you which stock is good or what to buy or sell. The KCE case above is an exercise in reading documents, not a view on the share price.

8. How to start tonight (and when to actually pay)

Do not pay yet. Two steps, in order.

Step 1 (tonight · free): open Boom Leverage Terminal and run a health check on the latest round in your book — type the risk you fear most in plain Thai ("ต้นทุนขึ้นจนมาร์จิ้นโดนบีบ" [input costs repriced faster than they could pass through, and the gap is sitting in gross margin] · "เลื่อนแผนลงทุน" [capex deferred — now find out how many times] · "ลูกค้ารายใหญ่หาย" [a key account walked and the revenue concentration is now exposed]) and read what management wrote themselves. Free, 100 credits/day, no card. Same full-market sweep as every paid tier, in every respect.

Step 2 (the day you actually need it): the day you hit an odd sentence in a fresh filing and the next question in your head is "hold on — have they said this before?" — that is the day to upgrade. Not before. Because on that day you are not buying software. You are buying 4 years (฿799) or 6 years (฿1,399) of magnification on the past to answer a question the free tier genuinely cannot.

Further reading:

• Which desks turn the Terminal into money — four real roles
• How semantic search differs from exact-match search, and why it matters
• Watch your book like a fund: upload a CSV/Excel watchlist and scan only your names
• Reading MD&A across years: why one company tells a different story every year (the PSL case)

Team/institutional (seats · Excel export · API): talk to us at the Enterprise page or contact@boomleverage.com

This content is for education and for accelerating research with citations. It is not investment advice · every price, quota and year window here is what the system actually runs as of the article's update date, and every quoted box is verbatim from a Form 56-1 that the listed company disclosed publicly through the SEC (check any box against the original) · no guarantee or forecast of security prices