Part 1
First, what kind of tools
are these?
Solar and Document Parse do different jobs. Once you understand the role of each, you can decide for yourself which one your work needs.
What this part covers
Now that you have your first response from Part 0, it is worth understanding how that response came about. Knowing the mechanics tells you what these tools are good for and what they are not.
What a language model is
How it reads and continues text, and what a token means.
What Solar is good at
Four kinds of work: summarizing, classifying, judging, and writing.
What Document Parse does
How it converts documents to text while preserving their structure.
How to combine the two
Most real automation is these two chained together in order.
1.1What a language model is
A language model works by reading what comes before and continuing the text. Because it has learned from an enormous amount of writing, what it produces reads naturally and fits the context.
So our job is to decide what it should continue from. The text you hand to the model is called a prompt.
The token
Models process text in pieces called tokens rather than letters or words. Roughly speaking, a token is a few characters in English and one or two characters in Korean.
Tokens matter for two reasons. First, usage is counted in tokens — the 50,000 tokens per minute limit from Part 0 refers to this. Second, there is a limit to how much a model can look at in one go.
| Term | Meaning |
|---|---|
| Prompt | The instructions and material you hand to the model |
| Token | The unit used to count text, and the basis for usage limits |
| Context | The total amount the model can look at in one request |
| Reasoning | The thinking a model works through before answering |
Solar does not store previous turns on the server. Each request is handled independently.
To go back and forth, you need to send the earlier questions and answers again. Keep in mind that the amount you send grows as the conversation gets longer.
A language model continues the text you hand it. Deciding what to hand it is the whole skill.
1.2What Solar is good at
Here are four kinds of work Solar handles well. It helps to see where your own work fits.
| Type | What it does | Examples |
|---|---|---|
| Summarize | Shortens long text | Pulling decisions out of meeting notes, condensing a paper's abstract |
| Classify | Sorts into categories you define | Routing enquiries to the right team, telling document types apart |
| Judge | Decides against a standard | Checking whether an application meets requirements, flagging amounts over a threshold |
| Write | Produces new text | Draft announcement emails, draft notices, draft teaching materials |
What suits it, and what to be careful with
Solar is strong at working with material you give it. Ask it about facts it does not know, and you may get an answer that sounds right but is not.
| Well suited | Be careful |
|---|---|
| "Pull out only the decisions from these meeting notes." | "What was our organization's budget last year?" |
| "Given the rules below, judge whether this application meets the requirements." | "Tell me what changed in the law this year." |
The difference is whether you supplied the material. Include the source with your request and the answers get far more reliable.
First, include the material. Put the document or the rules your judgment depends on into the prompt.
Second, state your criteria as numbers and conditions. "Flag it for rejection if the total is over 100,000 won" is far more stable than "please review this".
Third, specify the format. "As a table", "in three lines", or "start with either PASS or HOLD" gives you consistent results.
Solar is strong at summarizing, classifying, judging, and writing. Supply the material and state your criteria clearly.
1.3What Document Parse does
Document Parse converts files — PDFs, images, Korean HWP documents — into text that keeps its structure. It does not just pull out characters; it identifies headings, paragraphs, tables, and figures, then arranges them in reading order.
Why you need it
Solar works with text, so a document has to become text first. Scans and photographs are just images, so they cannot be read as they are. Document Parse bridges that gap.
What you can put in
| Item | Details |
|---|---|
| Supported formats | JPEG · PNG · BMP · PDF · TIFF · HEIC · DOCX · PPTX · XLSX · HWP · HWPX |
| Maximum file size | 50MB |
| Maximum pages | 100 pages (synchronous) · 1,000 pages (asynchronous) |
| Recognized scripts | Korean · Latin · numerals · Hanja · Katakana · Hiragana |
Being able to submit Korean HWP and HWPX files directly is particularly useful for work at Korean institutions.
Three processing modes
| Mode | When to use it |
|---|---|
| standard | Text-centred documents with simple tables. Enough for most cases |
| enhanced | Documents with complex tables, figures, or charts. Higher accuracy |
| auto | Chooses per page automatically. Handy when documents are mixed |
Higher resolution means better accuracy. Aim for a document width of at least 640 pixels.
The smallest text should be at least 2.5% of the image height. For a 640-pixel-tall image, that means at least 16 pixels.
What the result looks like
Results come back as HTML or Markdown. Tables stay tables and headings stay headings, which makes the text much easier for Solar to work with afterwards.
Each element also records which page and where on the page it came from — useful later when you need to check a value against the original.
Document Parse is the bridge between documents and language models. Preserving structure makes a real difference with tables.
1.4Putting the two together
Almost every automation you will build is a combination of the three steps below. Once you know the sequence, you can design for any kind of document.
Turn the document into text
Send the file to Document Parse and get back text that keeps its structure.
Pull out what you need
Ask Solar something like "find the applicant, amount, and date in this document". Fix the output format and you get the same shape every time.
Judge against your criteria
Compare the extracted values against your rules to reach a conclusion such as approve or hold. Stating criteria as numbers and conditions keeps results consistent.
Where this applies
| Work | What to extract | What to judge |
|---|---|---|
| Expense receipts | Merchant · date · total · line items | Whether it exceeds a limit, whether evidence is missing |
| Grant applications | Organization · project · amount requested · attachments | Whether all required documents are present |
| Literature review | Title · authors · method · key findings | Relevance to your research topic |
| Contract review | Term · amount · penalty clauses | Whether any clause is unfavourable or ambiguous |
The point of automation is not to remove people, but to narrow down what people need to check.
Rather than treating the output as final, keep a step where a person reviews borderline judgments and important values such as amounts.
Convert, extract, judge. Those three steps are the backbone of most document automation.
1.5[Activity] Break down your work
Pick one thing you do regularly and break it into the three steps above. What you write here carries straight into Parts 2 and 3.
# Breaking down my work
## The task I chose
- Name of the task:
- How often (per week / month):
- Time it takes per item:
## Step 1 · Turning documents into text
- File formats involved (PDF / image / HWP / other):
- How many at a time:
- Are they scans or photos:
## Step 2 · What to extract
- Field 1 (name / description):
- Field 2 (name / description):
- Field 3 (name / description):
## Step 3 · Criteria for judgment
- What conclusion is needed (e.g. approve / hold / reject):
- The criteria, written as numbers and conditions:
## Final check
- Values a person must always verify:
- What goes wrong if it is incorrect:Checkpoint
- The task you chose involves documents.
- You listed at least three fields to extract.
- Your criteria are numbers and conditions, not "review it".
- You identified at least one point where a person must check.
Start small. "Checking the amount and date on receipts" will get you results far faster than "all the document work in my department".
Part 1 key takeaways
- A language model continues the text you hand it. Deciding what to hand it is the skill.
- Solar is strong at summarizing, classifying, judging, and writing. Supply material and state criteria as numbers.
- Document Parse turns files into text that keeps its structure, including Korean HWP files.
- The backbone of document automation is convert → extract → judge.
- Automation exists to narrow what people check, not to remove them.
Where to go next
Time to build something. Pick whichever path suits you.
Prefer no code → Part 2
Connect blocks on screen. No development experience needed.
Prefer code → Part 3
Build it in Python for finer control.