User Guide
A short tour of what each part of CasePrepd does, with the bits that aren't obvious from the buttons alone.
For the CLI, batch processing, debug mode, and advanced settings, see the Power User Guide.
The Four Tabs
After you load documents and click Process Documents, the results pane on the right shows four tabs. Each one is doing a different job; switching between them is how you actually use the output.
- Document — The raw extracted text from whichever file you clicked in the left panel. Use this to confirm OCR worked, find a specific quote, or just read. Page-by-page navigation; Ctrl+F to search within a document.
- Vocabulary — The list of names, medical terms, legal phrases, and unusual words pulled from your documents. This is the main output for building a job dictionary. See the Vocabulary section below.
- Search — Ask questions in plain English, get answers with source citations. See Search Tips below for how to phrase questions.
- Key Excerpts — Auto-selected representative passages from your documents, grouped by topic. A starting point for getting oriented in a big case. See the Key Excerpts section.
The four tabs all work from the same processed batch of documents — you don't need to re-process to switch between them.
Vocabulary — What the Columns Mean
The Vocabulary table is where you'll spend the most time. Each row is a term CasePrepd thinks is worth showing you. The columns tell you why.
Quality Score
A 0–100 score combining several signals: how rare the word is in general English, whether it looks like a name or technical term, how it's used in context, and your past Keep/Skip ratings. Higher = more likely to be useful for your job dictionary. The default sort is by this column, descending.
# Docs
How many of the loaded documents the term appears in. A name that shows up in 6 out of 8 transcripts is probably a key witness. A name that appears in just one document might be a passing reference.
Occurrences
Raw count of how many times the term appears across all documents. Useful for sanity-checking: a name appearing 200 times in one transcript is almost certainly someone you need to know.
Found By
Which extraction algorithms picked up this term: NER (general-purpose name detection), MedicalNER (medical terms), RAKE, YAKE, and TopicRank (statistical keyword extractors), and BM25 (corpus-based rarity). Terms found by multiple algorithms are usually higher confidence.
Wondering why some name variants get collapsed into one entry and how the canonical spelling gets picked? See The Name Canonicalizer on the Built for Court Reporters page.
The Keep/Skip Thumbs — This Is Important
Every term row has a green thumbs-up (Keep) and a red thumbs-down (Skip). Use them. They're not just for filtering this one job — they train a model that gets better at flagging the kind of terms you care about over time.
A few patterns that pay off:
- Skip the obvious junk aggressively. "the", "said", "deposition", OCR garbage. The more junk you mark, the less of it you'll see next time.
- Keep names of people, places, and medical/technical terms — even ones that already look high-scored. It reinforces what's working.
- Click the same thumb again to clear it. Changed your mind? Click Keep a second time and the rating goes back to neutral.
The model starts contributing after about 30 ratings and keeps improving as you add more. Don't worry about being perfect early on — later ratings can override earlier ones.
Corpus
The app calls this feature a corpus — the Corpus... button in the top-left of the main window opens the manager. A corpus is a saved collection of past job documents (transcripts, exhibits, medical records, anything) that CasePrepd uses as a baseline for comparison. When you process a new job, terms that appear in your active corpus are de-emphasized — they're "things you already know" — while terms unique to the current job float to the top.
The point: legal boilerplate ("court", "objection", "deposition") is common in every transcript you'll ever read. You don't need it in your job dictionary. A corpus teaches the app what counts as background noise for your line of work.
Building One
Click Corpus... in the top-left of the app (callout 1 on the homepage screenshot), then + New Corpus. You can create as many corpora as you want — one for med-mal, one for criminal work, one for patent cases, etc. Each corpus is a folder of documents you add over time, and you pick which one is active in the Corpus: dropdown at the top of the main window.
Practical starting point: add 5–10 past transcripts from similar cases to a new corpus. That's usually enough for the effect to kick in. Adding more refines it; you don't need hundreds. The app shows a "5+ docs required for corpus functionality" hint when the active corpus is too small to help yet.
When It Helps Most
A corpus shines when your case has a lot of standard courtroom, medical, or technical boilerplate pulling attention from the things that actually matter. If your terms list is dominated by junk like "the witness testified," loading a corpus will fix that.
Semantic Search — How to Phrase Questions
The Search tab understands the meaning of what you ask, not just the keywords. That's powerful, but it means how you phrase the question matters more than you'd expect.
Good Phrasings
- "What injuries did the plaintiff describe?" — specific topic, full sentence.
- "When did the accident happen?" — asks for a specific fact.
- "What did Dr. Smith say about the MRI results?" — targets a person + a topic.
Less Effective
- "injury" — single keyword. Works, but you'll get a lot of loosely-related results. Try a full question instead.
- "Tell me everything about the case" — too broad. The search picks the single best passage, so vague questions get vague answers. Narrow it down.
When You Get No Match
If your question doesn't match anything in the loaded documents above a confidence threshold, the result reads: "No relevant excerpts found in documents." That's the app being honest about not finding anything, rather than picking the least-bad passage and showing it as an answer. Try rephrasing the question more specifically, or check that the relevant document is actually loaded.
The threshold isn't exposed in the Settings dialog, but power users
can adjust the retrieval_relevance_gate value in
%APPDATA%\CasePrepd\config\user_preferences.json
(default is 0.50). Lower it for weaker matches; raise it for
stricter "no match" behavior.
Key Excerpts — What to Expect
Key Excerpts pulls representative passages from across your documents and groups them. The idea: skim 8–15 passages and you've got a rough sense of what the case is about, who the players are, and what topics come up.
Important: this is not a summary. CasePrepd doesn't rewrite anything — the excerpts are verbatim from your documents. And not every excerpt will be a winner. Think of it as a sampler. Skim through, ignore the ones that aren't useful, and the good ones save you real reading time.
The excerpt count auto-scales with the document length —
roughly one excerpt per five pages, capped at 5–15 total. There's
no manual count setting; if you want more for a very large case, the
cap can be raised by editing user_preferences.json.
Where Things Live on Your Computer
CasePrepd stores its data and logs under your Windows user folder. Useful to know when something goes wrong or you want to back up your trained vocabulary preferences.
-
Vocabulary ratings (your Keep/Skip history):
%APPDATA%\CasePrepd\feedback\user_feedback.csv -
Saved corpora:
%APPDATA%\CasePrepd\corpus\ -
Settings:
%APPDATA%\CasePrepd\user_preferences.json -
Logs (if something crashed):
%APPDATA%\CasePrepd\logs\caseprepd.log
%APPDATA%\CasePrepd\logs\main_log_YYYYMMDD_HHMMSS.txt
To get there fast: open File Explorer, paste %APPDATA%\CasePrepd
into the address bar, hit Enter.
Common Issues
The app is slow on the very first run
Expected. The first time you process documents after installing, CasePrepd has to load several AI models from disk into memory. This can take 30–60 seconds. Subsequent runs reuse the loaded models and start almost instantly.
OCR is taking forever on a scanned PDF
OCR is genuinely slow — about 1–2 seconds per page on a typical machine. A 300-page scanned deposition can take 5–10 minutes. You can watch progress in the status bar. The Stop button cancels safely; partial results are kept.
The Vocabulary list has too much junk
Two things help: (1) Mark the junk with thumbs-down aggressively — the model learns fast. (2) Build a corpus from a few past transcripts so boilerplate gets de-emphasized automatically.
Search gave me a citation that doesn't seem right
If a real passage is returned but it isn't on point, the question probably needs to be more specific. If the search couldn't find anything confident enough to return, you'll see "No relevant excerpts found in documents." instead — try rephrasing the question more specifically, or check that the relevant document is loaded.
Windows SmartScreen blocked the installer
CasePrepd isn't code-signed (signing certificates cost hundreds of dollars a year for a free app). Click "More info" then "Run anyway" in the SmartScreen dialog. The full source is on GitHub if you want to inspect it before installing.
The app crashed or won't start
Check the crash log at %APPDATA%\CasePrepd\crash.log
and the main log at %APPDATA%\CasePrepd\logs\caseprepd.log.
If you can't figure it out, post the relevant log snippet to the
GitHub
Issues page — with logs, problems are usually fixable in
a day or two.
For Advanced Users
If you want to run CasePrepd from the command line (for batch jobs, scripting, or unattended processing), see the Power User Guide. It covers the CLI, every flag, debug mode, and the settings worth tweaking.