CampFire is an extractive contract-intelligence tool. Its models find and return text from your documents. They never write new text, and every result links back to its source. Every file you upload goes through the same stages before review:
It turns every page of an uploaded file into machine-readable text. It reads each page with OCR and keeps the position of every line on the page. Later stages use those positions to highlight results in the original document.
No. Scanned PDFs, image-only PDFs and native PDFs all go through Textualize. Native text is picked up directly; scanned pages are recognized.
Highlights are drawn from the OCR line positions. Skewed scans, low resolution or stamps over text can shift the recognized position by a few pixels. The extracted text is still correct.
The models can only find what OCR can read. Faint, handwritten or heavily marked-up pages give weaker text, so clauses on those pages are more likely to be missed. Rescanning at 300 DPI or higher gives the best results.
Yes. Every textualized page counts toward your plan's monthly pages (see Subscriptions).
A file is what you upload. A document is one logical agreement inside it. A single PDF can hold a master agreement, three amendments and a set of exhibits. Unitize splits it into separate documents and lists each one in the Work tab.
It looks for page-break signals in the textualized pages:
These signals are combined into a suggested set of document boundaries.
The boundaries are suggestions. You can merge or split documents in the review interface, and downstream results follow the corrected boundaries.
No. A file that has already been unitized is not split again. A repeat upload of an identical file is caught by near-duplicate analysis (see Suppression).
It breaks each document into its smallest reviewable pieces (atoms) and runs the contract models over them. Atoms include sentences, clauses, named parties and entities such as dates and amounts.
For each document, Atomize produces:
Each result carries its character range and page position, so the reviewer can jump straight to it in the source.
Small pieces keep every answer tied to a specific passage. A reviewer can check any result in seconds, and nothing is paraphrased or invented.
Yes, when a newer model version is released or a project is reprocessed. Each run is recorded with its model name and version, so results from different versions don't mix.
The file moves through the pipeline automatically. Each stage starts when the previous one finishes:
Results show up in the Work tab as each document finishes. You don't have to wait for the whole upload.
It depends on page count and scan quality. Large scanned files take longest, because OCR is the slowest step.
A failed step retries automatically, up to five attempts. A file that still fails is set aside for the CampFire team to inspect, and the rest of your upload carries on. Contact support if a file stays stuck.
Yes. Archives are unpacked and each file inside is processed on its own. Storage counts both the original archive and the unpacked copies.
Suppression hides a result, or a whole document, from review views and exports without deleting it. It is a one-click control (the eye icon) on every extracted value.
When the model flagged a passage that is not what the column asks for. Suppressed results drop out of the Grid, filters and CSV exports, so the deliverable shows only confirmed findings.
No. Suppression is reversible. Show hidden items again and un-suppress anything hidden by mistake.
Near-duplicate analysis compares each new file to files already in the project, by page count, length and text similarity. Likely duplicates are flagged so a reviewer can suppress them, instead of reviewing the same contract twice.
A suppression is a reviewer's judgment that the result is wrong. It can feed the training data for Annotation Models (see Annotation Models).
A single-document view that shows everything CampFire found in one agreement. It lists the document type, parties, key dates and amounts, and every clause the models flagged, each linked to its highlighted passage.
Use the Grid to compare one question across many documents. Use Doc X-Ray to understand one document in depth, such as an unusual side letter or a key customer contract.
Yes. Suppress a wrong result, or annotate a missed span, directly in the document. The change shows up in the Grid and in exports.
The main review table in the Work tab. Each row is a document, and each column is a clause model or question. Each cell holds what CampFire found for that document.
Use the Columns menu to show or hide models and questions. Filters and search narrow the rows, for example to every document with a Change of Control clause.
Click it to open the source document at the highlighted span. Confirm it, suppress it, or annotate a better span.
Export to CSV, either all rows or only the filtered set. Suppressed results are left out of the export.
Every plan includes all 40 clause models, the full review interface and CSV export.
| Plan | Price / month | Users | Projects | Pages / month | Storage | Extra page | Extra user / month |
|---|---|---|---|---|---|---|---|
| Solo | $50 | 1 | 1 | 200 | 500 MB | $0.30 | $20 |
| Team | $400 | 20 | 25 | 2,000 | 5 GB | $0.25 | $15 |
| Company | $1,000 | 100 | Unlimited | 5,000 | 12 GB | $0.15 | $10 |
| Enterprise | Custom | Custom | Unlimited | Custom | Custom | Custom | Custom |
Team and above add project archiving. Company adds management reports. Enterprise adds SSO, custom integrations and a dedicated environment.
By the high-water mark during the billing period. Storage includes the originals of compressed files and their unpacked copies.
Processing continues, and extra pages are billed at your plan's overage rate.
Yes, as long as each organization has a free seat.
A review item you want answered for every document, such as "Can this agreement be assigned without consent?" Each Question becomes a column in the Grid.
No. CampFire is extractive. The answer to a Question is the passage, or passages, in each document that addresses it, linked to its source. Your reviewers draw the conclusion.
The 40 built-in clause models cover standard diligence topics. A Question lets you track a topic specific to your deal, and its answers can be annotated and turned into a model over time.
A label on an exact range of characters in a document, for example marking the sentence that grants a termination right as "Termination for Convenience". You can tag whole documents or individual spans.
A span points to the exact words that matter. That makes results quick to check, keeps highlights accurate, and gives the models precise examples to learn from.
Open the document, select the text and choose the label. The new span appears in the Grid and in exports like any model result.
Yes. Each annotation records whether a model or a reviewer made it, and which model version.
The models that produce CampFire's results. They cover clause classification, sentence categories, span extraction, parties and entities, and document type. Every plan includes 40 pre-built clause models, ready to run.
From expert span annotations on real agreements. Reviewer corrections, both suppressions and added spans, show where a model is wrong and can feed the next training round.
New versions are released centrally and documents can be reprocessed with them. Results are stored with the model name and version, so you always know which version produced a finding.
Talk to our team about custom models for Enterprise plans.
Recall is the share of real clauses the model finds. Precision is the share of flagged results that are real clauses. Pushing one up usually pulls the other down.
In other words, favoring recall means casting a wider net. This allows for slightly more false positives but greatly decreases the chance of CampFire missing a clause.
In diligence, a miss costs far more than a false alarm. A missed change-of-control clause can derail a deal or create liability. A false positive costs a reviewer a few seconds and one click to suppress. So the models are tuned to flag anything that could plausibly be a match.
Expect some results that turn out not to be relevant. That is the system working as designed. Review each flagged result in context and suppress the ones that don't fit. The final export then has both high recall and high precision.
It would look cleaner but be less safe. Anything the model dropped would never reach a reviewer, so nobody would know it was missed. With a recall bias, the human decision covers every borderline case.