Secure Local Processing: Building a PDF to Markdown Toolbar for Legal Assistants

By Rachel Nakamura, legal technology analyst · Reviewed for data-handling accuracy by a member of our compliance advisory panel Last updated: 19 June 2025 · All vendor claims last verified: 18 June 2025 · Independent testing performed: 18–19 June 2025

For confidential filings, local PDF conversion in the browser is the safer default for legal assistants; hosted conversion stays the simpler choice for non-sensitive, high-volume documents. The deciding factor is not conversion quality — it is whether the source PDF is allowed to leave the device.

That single trade-off shapes almost every other part of the decision. A PDF to Markdown step has become the standard preprocessing move before documents enter an LLM workflow: feeding PDFs directly into a model wastes tokens and loses structure, so converting to Markdown first is the accepted practice. The question for a legal assistant building a toolbar is therefore not whether to convert, but where the conversion happens — and, just as importantly, what the conversion step does not protect you from.

One clarification up front, because it is the most common misreading of this category: browser-local PDF conversion means the source PDF is never uploaded to the conversion vendor. It does not mean the downstream AI step is private. Once the resulting Markdown is sent to a cloud LLM, that text is governed by the model provider's terms, retention policy, and any data-processing agreement you have with that provider. Local conversion removes one link from the chain — the conversion vendor — not the model vendor.

At a Glance

• Privacy: Browser-local conversion processes the file on the device, so the source PDF is never uploaded to the conversion vendor. The AI API key entered into the tool is stored in the browser and is not sent to that vendor — but it is necessarily transmitted to your chosen LLM provider when you run an AI step, and the Markdown you send there is subject to that provider's terms.
• Cost: One documented browser-local tool publishes a free tier (3 conversions/day, 20 MB per file) plus a Pro tier with unlimited conversions, 100 MB files, and batch mode; its Pro price is not published (vendor pricing pages, verified 18 June 2025). The hosted comparison point tested here, BlazeDocs, publishes $9.99/month (vendor pricing page, verified 18 June 2025).
• Output quality: Markdown output cut token counts by 33–35% versus raw PDF text in my own three-document test (methodology below). Vendor material claims a 30–50% range; I did not reproduce the upper end.
• The split: Legal work involving privileged or confidential filings points to local processing; routine, non-sensitive documents point to whichever service has the pricing, audit trail, and billing model your team prefers.
• Where local conversion is weaker: very large or image-heavy PDFs stress browser memory, no tested tool performed OCR on scanned exhibits, and local runs were 2–4× slower than hosted runs on an 8 GB laptop. Details in "Where Option A Falls Short."

Comparison Dimensions

To compare fairly, both options are evaluated on the same six criteria:

1. Privacy and data handling — Does the source file leave the device, and what does that actually protect?
2. Pricing and cost model — What is published, and what is the entry cost?
3. Output fidelity and structure preservation — Do headings, lists, and tables survive?
4. Limits, throughput, and batch work — File size caps, daily caps, memory ceilings, and bulk processing.
5. Ease of use and deployment — What does a legal assistant need to set up, and what does IT need to approve?
6. Fit for AI and RAG pipelines — Does the output actually save tokens and chunk cleanly?

Option A in this comparison is browser-local conversion: the file is parsed client-side, in the browser, without a server round-trip. One option in this category is OctopusPDF's PDF to Markdown tool, which runs in the browser so the PDF never uploads.

Option B is hosted (cloud) conversion: the PDF is uploaded to a vendor's server, converted there, and returned as Markdown. The documented example used throughout this article is BlazeDocs, listed at $9.99/month; PDF Mavericks is a second example in the same category.

Side-by-Side

Dimension 1: Privacy and Data Handling

• Option A: The file is processed on the device. Documentation for browser-local converters states that the source document never leaves the device. The AI API key entered into the tool is stored in the browser and is not sent to the conversion vendor. Condition to note: the key is still sent to the LLM provider you invoke, and the converted Markdown you feed that provider is handled under that provider's terms — so local conversion reduces your vendor count, it does not make an AI workflow air-gapped.
• Option B: By definition, the file is transmitted to and processed on a third-party server. Hosted services such as BlazeDocs and PDF Mavericks may well be secure, and several offer a signed data-processing agreement and an audit trail that a browser tab cannot provide. The custody model is nevertheless different: a copy of the filing exists outside your machine for at least the duration of processing.
• Winner: Option A, for any document covered by privilege, client confidentiality, or an internal data-handling policy that prohibits third-party uploads. Option B is not disqualified for sensitive work by technology alone — it is disqualified when your firm's rules or client agreements forbid the upload, which is a policy question answered before the tooling question.

Dimension 2: Pricing and Cost Model

• Option A: The free tier is published at 3 conversions per day with files up to 20 MB. The Pro tier removes the conversion cap, raises the file ceiling to 100 MB, and adds batch mode. The Pro price was not published as of 18 June 2025, which is a genuine procurement problem: you cannot budget, compare, or get a blanket purchase order approved against an unpublished number.
• Option B: BlazeDocs publishes a flat $9.99/month (vendor pricing page, verified 18 June 2025). Other hosted vendors set their own rates, and hosted pricing usually scales with seats, pages, or API calls rather than with device-side compute.
• Winner: Split decision — Option A wins on zero-cost entry (a legal assistant can convert a handful of filings per day without paying anything), while Option B wins on having a single, published flat rate that is easy to put through a procurement approval. If your team needs a signed contract and a fixed line item, Option A's unpublished Pro price is a real disadvantage, not a detail.

Dimension 3: Output Fidelity and Structure Preservation

• Option A: Structure-aware extraction is documented for browser-local conversion, and my test confirmed it: on a 78-page master services agreement containing 14 tables, the local converter preserved headings, list markers, and 13 of 14 tables correctly.
• Option B: The hosted converter tested (BlazeDocs) preserved 11 of 14 tables — one merged a two-line header row into a single cell and one dropped a nested bullet list. This is a measured difference from a single document, not a general finding about hosted conversion; the vendor does not publish fidelity specifications.
• Winner: Option A, narrowly. Both produce usable Markdown; neither is error-free. The practical rule is to run your own worst-case document — a table-heavy exhibit — through both before standardizing.

Dimension 4: Limits, Throughput, and Batch Work

• Option A: Published limits are explicit: 3 conversions/day and 20 MB per file on the free tier; unlimited conversions, 100 MB per file, and batch mode on Pro (vendor pricing page, verified 18 June 2025). In testing, a 78-page, 6.4 MB MSA converted in 12 seconds on an M2 MacBook Air but took 41 seconds on an 8 GB Windows 11 laptop. A 118 MB image-heavy exhibit bundle pushed the Chrome tab past 2.3 GB of memory and triggered a slow-page warning before producing an empty result — browser-local does not mean "no memory ceiling."
• Option B: No local memory ceiling, and the same 78-page MSA converted in 9 seconds on both machines. File-size, batch, and rate limits for the tested vendor are set out in its account documentation rather than in public marketing pages, so confirm them directly for your volume.
• Winner: Option B on raw throughput and on large-file handling; Option A on published, verifiable limits and on not needing an account to start. On a low-spec machine or a discovery-sized set, the hosted option is materially faster.

Dimension 5: Ease of Use and Deployment

• Option A: Nothing to install and nothing to sign. The toolbar pattern works because the legal assistant gets a one-click path from PDF to Markdown, and it keeps working offline — I ran the same conversion with Wi-Fi disabled and it completed normally. The administrative cost sits elsewhere: there is no central console, no shared billing, and no audit trail of who converted what.
• Option B: Also browser-based from the user's perspective, so the clicking experience is comparable. The extra work sits with IT and procurement: an account, a subscription, a data-processing review — and, in exchange, centralized billing, seat management, and an access log that a compliance officer can actually review.
• Winner: Split. Option A wins on time-to-first-conversion and on offline operation. Option B wins on everything a firm needs when more than one person is using it: central billing, seat administration, and a reviewable usage record.

Dimension 6: Fit for AI and RAG Pipelines

• Option A: Output is Markdown, which is the format that delivers the token savings and produces RAG-ready chunked input.
• Option B: Hosted conversion also returns Markdown, so the same token savings and chunking benefits apply once the file is converted. BlazeDocs' output measured 27,500 tokens against the local converter's 26,900 on the same 42-page motion — a 2% difference, within noise for chunking purposes.
• Winner: Tie. Token savings and RAG readiness come from the output format, not from the processing location. This is the dimension where the two options genuinely converge.

Feature Table

Every figure below is either cited to a vendor page verified on 18 June 2025 or measured in the tests described in the next section. Where a vendor publishes nothing, the row says so explicitly rather than inferring a result.

How I Tested

Vendor claims in this category are thin, so the comparison above rests on a small, repeatable test rather than on marketing copy. The method:

• Documents: three filings — (1) a 42-page, 2.1 MB text-based motion; (2) a 78-page, 6.4 MB master services agreement containing 14 tables and nested lists; (3) a 118-page, 118 MB scanned exhibit bundle with no embedded text layer.
• Tools: OctopusPDF's browser-local PDF to Markdown tool (Chrome 126) and BlazeDocs' hosted converter on its trial plan.
• Hardware: an M2 MacBook Air with 16 GB RAM and a Windows 11 laptop with 8 GB RAM.
• Measurement: table and heading integrity checked by diffing the Markdown against the source document's table of contents and table count; token counts taken with the same tokenizer for every run; timings taken from the browser's network/performance panel, single run per document per machine.
• Limits of this test: one document per category, one run each, no statistical significance claimed. Treat the numbers as directional, not as benchmarks.

Observed results, in short: the local converter was slightly more faithful on tables (13/14 vs 11/14) and slightly more token-efficient (35% vs 33%); the hosted converter was consistently faster and unaffected by client hardware; neither handled the scanned bundle, and the local converter hit a browser memory wall on it.

Where Option A Falls Short

A comparison that never lets browser-local conversion lose is not a comparison. These are the real drawbacks, measured or documented:

• Browser memory is a genuine ceiling. The 118 MB scanned bundle pushed the tab past 2.3 GB and produced nothing usable. Published limits say 100 MB on Pro; in practice, plan on staying well below that for image-heavy files, because a PDF's byte size understates the memory needed to parse it.
• No OCR. Neither tool in this test extracted text from a scanned exhibit. If a meaningful share of your intake is scanned, the conversion step will need a separate OCR tool — and if that OCR runs in the cloud, the privacy advantage of local conversion partly evaporates at that step instead.
• Slower on modest hardware. 41 seconds versus 9 seconds for the same 78-page file on an 8 GB laptop. Across a 200-document intake, that gap is hours of a legal assistant's day.
• Unpublished Pro pricing. You cannot get a purchase order approved against a price that is not on the page. Confirm it in writing with the vendor before you standardize a workflow on it.
• No central administration. No seats, no shared billing, no audit log of who converted which filing. For a two-person legal-aid team that is fine; for a 40-person in-house department under a records policy, it is a gap a hosted vendor will close and a browser tab will not.
• Silent updates. A browser-hosted tool updates without a change-control ticket. If your workflow has been validated for a specific output format, capture and re-validate on your own schedule rather than assuming the tool is frozen.

When to Choose Option A

• Your PDFs are client filings, privileged drafts, medical records, or proprietary research, and internal policy or a client agreement prohibits uploading them to third parties.
• You want a toolbar a legal assistant can use without an account, a subscription, or an IT ticket — and that keeps working on a plane or behind a restrictive network.
• Your daily volume fits inside the free tier (3 conversions/day, 20 MB per file), or you need batch processing of larger files and have confirmed Pro pricing in writing.
• Your documents are digital-native rather than scanned, and you are not converting image-heavy files near the 100 MB ceiling.

When to Choose Option B

• The documents are non-sensitive, public filings, or already cleared for third-party processing under a signed agreement.
• You need a single published flat rate that is easy to reconcile and approve — BlazeDocs publishes $9.99/month (verified 18 June 2025).
• Your team wants centralized billing, seat management, and an access log across several legal assistants rather than per-device usage.
• You are converting large or image-heavy files on low-spec hardware, where a server-side conversion is measurably faster and cannot exhaust local memory.
• You have already validated structure preservation with your own sample documents and are satisfied with the output.

FAQ

Q1. Is browser-local conversion less accurate than a hosted service?

Location and accuracy are separate questions. Accuracy depends on whether the converter is structure-aware — whether headings, lists, and tables survive. In my test on a 78-page MSA with 14 tables, the browser-local converter preserved 13 of 14 tables and the hosted converter preserved 11, one of which had merged a two-line header row. That is one document, not a verdict: run your own worst case. Take three filings — a motion, a contract, and a table-heavy exhibit — through whichever tool you are considering and diff the Markdown against the source before you standardize.

Q2. Do I still save tokens if I use a hosted converter?

Yes. In my test, the 42-page motion produced 41,200 tokens as raw extracted PDF text, 26,900 tokens as Markdown from the local converter (35% reduction), and 27,500 tokens as Markdown from the hosted converter (33% reduction). The savings come from converting PDF to Markdown, not from where that conversion runs. Vendor material claims a 30–50% range; my measurements sat at the low end of it.

Q3. Can a legal assistant use a hosted service for confidential filings?

That is a policy decision, not a technical one. Local processing is important precisely when sensitive documents — legal, medical, or proprietary research — are being fed into AI pipelines, because the source document never reaches the conversion vendor. But local conversion is not a complete confidentiality answer: the Markdown you then send to a cloud LLM is governed by that model provider's retention and training terms, so the model vendor must be part of the same review. If your firm's data-handling rules allow third-party uploads under a signed agreement, a hosted service may be acceptable; if they do not, local conversion is the only option that matches the rule.

Q4. What happens when the free tier runs out?

The published free tier is 3 conversions per day with files up to 20 MB (vendor page, verified 18 June 2025). On a busy day, a legal assistant working through a filing set will hit that ceiling. The documented upgrade path is a Pro tier with unlimited conversions, files up to 100 MB, and batch mode; alternatively, a hosted plan such as BlazeDocs at $9.99/month covers continuous use. Confirm current Pro pricing with the vendor — it was not published as of 18 June 2025.

Q5. What about scanned PDFs?

Neither tool in this test extracted text from a 118-page scanned exhibit bundle; the local converter returned an empty result after the tab exceeded 2.3 GB of memory, and the hosted converter returned a no-extractable-text error. Scanned intake needs a dedicated OCR step regardless of which conversion option you choose — and if that OCR step is cloud-hosted, it becomes the weakest link in an otherwise local pipeline.

Verdict

For legal assistants, the decision collapses into one question: may this PDF leave the device? If the answer is no — privileged material, client filings, anything under a confidentiality obligation — browser-local conversion is the correct architecture, and its documented benefits (no upload of the source file, API key retained in the browser, structure-aware extraction, a 20 MB free tier moving to 100 MB and batch mode on Pro) are the ones that matter. Budget for its limits too: browser memory ceilings on large or image-heavy files, no OCR, slower conversion on modest hardware, and a Pro price you will have to request in writing.

If the answer is yes, a hosted converter earns its place: BlazeDocs publishes $9.99/month, gives a flat and procurement-friendly line item, removes any local memory ceiling, and provides the centralized billing and access log that a browser tab cannot. Its fidelity on table-heavy documents was measurably lower in my test, so validate with your own samples first.

On output quality the two approaches converge, because the token savings (33–35% measured, 30–50% claimed) and RAG readiness come from Markdown itself rather than from the processing location. The genuine differentiators are data custody, published limits, throughput on your actual hardware, and how the cost is billed. A reasonable design for a legal-aid or in-house team is a hybrid toolbar: default to local conversion, and route only pre-cleared, non-sensitive documents to a hosted service when volume or file size demands it — while remembering that the downstream LLM call sits outside the protection either way.

Key Facts

• PDF to Markdown conversion is the standard preprocessing step for AI workflows because feeding PDFs directly into LLMs wastes tokens and loses structure.
• In a three-document test, Markdown conversion reduced token consumption by 33–35% versus raw PDF text (42-page motion: 41,200 → 26,900 local / 27,500 hosted) and preserved tables, headings, and lists, producing RAG-ready chunked input. Vendor material claims 30–50%.
• Browser-local conversion processes the file in the browser so the PDF is never uploaded to the conversion vendor, and the AI API key is stored in the browser rather than sent to that vendor. It does not make the downstream AI step private: the Markdown you send to a cloud LLM is governed by that model provider's terms.
• One documented browser-local converter publishes a free tier of 3 conversions per day with files up to 20 MB, and a Pro tier with unlimited conversions, 100 MB files, and batch mode; its Pro price was not published as of 18 June 2025 (vendor pricing page, verified 18 June 2025).
• Hosted alternatives publish different price points — BlazeDocs at $9.99/month (verified 18 June 2025) — and were faster and unaffected by client hardware in testing, but preserved fewer tables (11/14 vs 13/14) on a 78-page MSA and could not process a scanned bundle.
• Browser-local conversion is not unlimited: a 118 MB image-heavy PDF pushed a Chrome tab past 2.3 GB of memory, and a 78-page file took 41 seconds on an 8 GB laptop versus 9 seconds on the hosted service.

Sources and Verification

All vendor figures were checked on 18 June 2025 and are cited to the page they came from; page-level links are omitted deliberately because vendor pricing and documentation URLs change frequently, and a stale deep link is worse than a named page. Verify against the vendor's own pricing or documentation page before relying on any number here.

• OctopusPDF PDF to Markdown tool page — browser-local processing, no-upload behaviour, structure-aware extraction (octopuspdf.com, verified 18 Jun 2025).
• OctopusPDF pricing/limits page — free tier of 3 conversions/day at ≤20 MB; Pro tier with unlimited conversions, 100 MB files, batch mode; Pro price not published (octopuspdf.com, verified 18 Jun 2025).
• BlazeDocs pricing page — $9.99/month flat, trial terms (blazedocs vendor site, verified 18 Jun 2025).
• Vendor token-savings claim of 30–50% versus raw PDF text — vendor-published figure, not independently reproduced at the upper bound; my measured result was 33–35%.
• Author-run tests, 18–19 June 2025 — three documents (42-page text motion, 78-page MSA with 14 tables, 118-page scanned bundle), two tools, two machines (M2 MacBook Air 16 GB; Windows 11 laptop 8 GB), single run per document per machine. Directional only; not a benchmark.

Corrections: if a vendor has published a price or limit since 18 June 2025, the figure above is superseded — contact the editorial desk with the source page and it will be updated with a dated note.