Essential cookies keep Instavar working. Optional analytics help us understand how the site is used. Cookie Policy
Manage Cookie Preferences
Service reliability telemetry, including Sentry error monitoring and Vercel Speed Insights, stays enabled so we can secure the product and diagnose failures.
Token Saver with Codex: A Smaller, Safer Way to Read Long PDFs · Instavar
Token Saver sends Codex only the pages that matter.
We tested Token Saver on long brand guides, research reports, and similar PDFs. It searches locally and gives Codex a few page-cited passages instead of every page.
88.9% less PDF text
Across five public PDFs, every one of our 15 test questions found an accepted source page in the first five search results.
Use it for long, text-heavy PDFs and focused questions. Read short documents directly, and inspect scans, diagrams, and layout visually. This test measured PDF text, not total cost or answer quality.
What this changes for a video team
A video project often starts with documents rather than footage. A team may need
to find the approved tone in a brand guide, a required disclaimer in a legal
brief, or the target audience in a campaign plan.
Codex does not need every page to answer each of those questions. Token Saver
keeps the full PDF outside the conversation, searches it on the local machine,
and gives Codex a small set of passages with page numbers.
This has three practical benefits:
More focused answers. Less unrelated text competes with the instructions
and evidence needed for the current question.
Easier review. Page references help a person check the source before a
production decision is accepted.
Reuse across questions. A long document can be indexed once and searched
again when the team asks a follow-up question.
These are operational benefits, not proven financial savings. We have not yet
shown that the complete Codex session costs less, finishes faster, or produces a
better final answer.
What the test showed
We tested a reviewed version of the open-source
Token Saver repository on five
public PDFs from official US government sources:
Copyright Basics
NIST AI Risk Management Framework 1.0
the Dobbs opinion
NIST SP 800-53 Revision 5
NASA Systems Engineering Handbook
The documents ranged from 10 to 492 pages and from about 5,184 to 332,621
extracted tokens, a rough measure of text size. We wrote 15 questions whose
answers were tied to known source pages. The NASA handbook was held back until a
later check so that it could test whether the approach worked on a new document.
Search method
Accepted page found in the top five
Exact-word search
13 of 15 questions
Exact words plus related meaning
15 of 15 questions
Across the 15 questions, Token Saver returned 88.9% less PDF text on average
than sending all extracted text. On the three largest documents, it returned
about 98.0% to 99.2% less.
This supports one limited finding:
In this test, Codex received much less PDF text while an accepted source page
still appeared in the first five search results for every question.
It does not show that every answer was correct. It does not show that the same
result will hold for any PDF or question. It also does not show lower total
Codex usage, cost, or time.
Less PDF text is not the same as a cheaper agent
The PDF is only one part of a Codex session. The system may also load project
instructions, conversation history, tool descriptions, earlier results, and
the model's own response.
Our first comparisons did not keep those other conditions equal. One control
did not answer the same PDF question. Another used a different set of tools and
permissions. The Token Saver run returned less PDF text, but its complete record
was larger because more supporting machinery was loaded.
That observation does not show that Token Saver caused higher use. It shows why
the comparison was not fair.
A fair test must keep the model, question, tools, permissions, and workflow the
same. It should change only one thing: whether Codex receives the complete PDF
text or the smaller set of retrieved passages. Building the index for the first
time should also be measured separately from reusing an existing index.
Until that test is complete, the safe conclusion is simple:
Token Saver reduced the PDF text given to Codex in our test. Its effect on the
cost and speed of the whole workflow is still unknown.
Search results are clues, not proof
Document search can return a passage that sounds relevant without supporting
the claim in the question.
We tested this by asking 10 questions that the selected documents did not
answer. The original Token Saver server still returned passages for all 10.
Simple checks rejected many obvious mismatches, but some false questions used
the same names and terms as the document and still looked convincing.
This is why a similarity score cannot be treated as truth. A court opinion may
mention the right law and person but quote a dissent rather than the majority. A
brand guide may mention a color without approving it for the requested use.
Our Codex workflow therefore follows five steps:
Register the PDF.
Ask one focused question.
Read the returned passages.
Check whether they cover the important parts of the question.
Answer with page references and state what remains uncertain.
The checks can reveal missing terms or weak coverage. They cannot prove that a
passage is complete, correctly attributed, legally binding, or sufficient for a
high-risk decision. A person or agent still has to read the evidence.
Choose the right way to read each PDF
The goal is to notice every PDF and choose the right reading method. The goal is
not to force every PDF through Token Saver.
PDF and task
Best starting method
Long, text-heavy PDF with repeated focused questions
Token Saver search
Short, text-only PDF
Read it directly
Scanned or image-only PDF
Read the text with OCR and inspect the pages
Diagram, typography, or layout review
Inspect the full pages visually
Whole-document summary
Read every section with a completeness check
Proving that something is absent
Search broadly and state exactly what was checked
We tested the difficult edge with six fresh Codex runs about a short, scanned,
diagram-heavy brochure. All six chose visual inspection first. None made Token
Saver the main method. This supports the table for that repeated task, but it is
not a general success rate for every model, machine, or document type.
Where it fits in Instavar
Token Saver can help during planning. It can find textual requirements such as:
the approved tone
the target audience
a required disclaimer
a prohibited phrase
a campaign objective
Those findings can become inputs to a script or production brief. They cannot
approve the finished video.
A page reference does not prove that:
an image, voice, or clip is licensed for the customer
the final composition follows the visual brand rules
a crop is safe for a vertical platform
captions and narration match the final timeline
the rendered video passed picture and sound review
the selected social account is authorized
a person approved the final publish action
The Codex plugin is therefore a local research and operator tool. It is not a
production feature inside Instavar Studio. A production version would also need
customer separation, source versions, access rules, index refresh and deletion,
and the existing render, review, and publishing controls.
Privacy: what stays local and what may leave the machine
The full PDF can stay on the local machine while selected passages are sent to a
cloud-hosted model. Those are different claims.
Local text extraction and local search reduce how much document text is sent to
the model provider. They do not guarantee that no document content leaves the
machine. The saved search index also contains document text and numerical search
data, so it needs the same care, retention period, and deletion rules as the
source PDF.
Folder allowlists, file checks, size limits, and source fingerprints reduce
risk. They do not replace the user's permission or the controls needed to keep
one customer's information separate from another customer's information.
We used only public, non-private PDFs in this evaluation.
How the Codex integration works
The main idea is simple:
PDF on the local machine
-> extract and divide its text into small passages
-> search for exact words and related meaning
-> remove repeated results
-> send a few page-cited passages to Codex
We packaged the integration as a Codex plugin. A plugin is the container that
groups the parts needed to install and use the feature. Our package contains:
a short instruction that tells Codex when to use PDF search
a prompt check that notices visible PDF references
a safety check before supported local PDF-reading commands run
a local server that registers, searches, and checks PDF passages
OpenAI's documentation explains the wider
Codex plugin and
hook systems.
The prompt check runs whenever a prompt is submitted, but it does not add text
to every conversation. It inspects the prompt and exits silently when there is
no visible PDF signal. When a PDF is involved, it adds the PDF reading rules.
In 100 local runs using ordinary prompts with no PDF, the check added no text to
the model's prompt. Its middle run time was 27.19 milliseconds on this Mac.
Turning on optional measurements added about 0.65 milliseconds to that middle
value.
This timing applies only to the tested Mac, script, and prompt set. It does not
prove the same speed on other machines or under heavy parallel use. Local hooks
also do not cover every possible hosted tool path, so they are a useful routing
check rather than a complete security wall.
The decision and the next test
We kept Token Saver as a selective local PDF reader for Codex.
Use it when the document is long and text-heavy, the question is focused, and
several questions can reuse the same index. Avoid making it the default when the
document is short, scanned, visual, or must be read in full.
The broader lesson is:
Sending less text helps only when the smaller selection still contains
everything needed for the decision.
A large reduction has negative value if it removes an exception, diagram,
speaker, date, or approval rule.
The next useful comparison will keep the full Codex environment fixed and
change only how the PDF text is supplied. It should measure total usage, time,
answer support, correction work, and human review. We also need new tests with
scanned, multilingual, table-heavy, damaged, oversized, and multi-document
cases.
For now, the supported conclusion remains narrow: local search can greatly
reduce the PDF text given to Codex for large text documents and focused
questions. Whether that improves the complete agent workflow still needs to be
measured.
If your video workflow begins with long research, brand, or compliance
documents, decide what evidence each production decision needs before choosing
how the documents will be searched.
Download the printable version
. Last updated 1 Aug 2026. This article reports one reviewed Token Saver version,
five public PDFs, 15 written questions, one macOS Codex environment, and follow-up
routing checks. Do not apply the results more broadly without further testing.