Evidra
Sign in
Product

Four modules, one evidence chain

Knowledge goes in once. Every questionnaire after that is drafted from it, reviewed by the people who own the answers and exported in the buyer's own format with the evidence attached.

Module 1

Knowledge library

Upload the documents your team already trusts. Each one carries a category, an owner, an effective date and an optional expiry date, so an answer drafted from an expired policy is visible as such.

  • Categories: Policy, Product, Security, Legal, Past response and Certification.
  • Ingestion status per document: Uploaded, Processing, Indexed or Failed, with the error shown when a file cannot be read.
  • Page counts and chunk counts after indexing, so you can see what the drafting step can reach.
  • A document viewer that opens at the cited page from any evidence passage.
EvidraKnowledge library, 4 documents
DocumentCategoryStatusPages
Evidra acceptable use policyLegalIndexed1
Evidra privacy policyLegalIndexed2
Evidra security overviewSecurityIndexed3
Evidra terms of serviceLegalIndexed2
Module 2

Questionnaire intake

Upload the questionnaire as you received it: Excel, CSV, Word or PDF. Extraction preserves sections, numbering and answer format constraints such as yes or no fields and character limits.

  • Spreadsheet sources keep the sheet, row and answer column, so the export writes back into the right cell.
  • Text sources keep the paragraph position so the export rebuilds the document in order.
  • Every extracted question can be corrected, merged or removed before drafting starts.
EvidraIntake, vendor-security-assessment.xlsx, 13 questions
Sheet: Security assessment. Questions in column B, yes or no to column C, answers to column D.
No.SectionQuestionFormatLimit
2.1Data protectionIs customer data encrypted in transit and at rest?Yes or no with detailNone
2.2Data protectionWhere is customer data stored and processed?Free textNone
2.3Data protectionList the subprocessors that process customer data and the purpose of each.Free textNone
2.4Data protectionIs customer data used to train or fine tune AI models?Yes or no with detailNone
2.5Data protectionHow is customer data separated from other customers' data?Free textNone
Module 3

Evidence first drafting

Drafting runs as a background job with progress per question. Each question is matched against your passages, the supporting passages are attached, and the answer is written strictly from them.

  1. 01

    Company knowledge

    Past responses, policies, product documentation and compliance records, uploaded once.

  2. 02

    Ingestion

    Each document is read page by page and split into passages that keep their page numbers.

  3. 03

    Retrieval

    Every question is matched against your passages only, never against the open web.

  4. 04

    Evidence

    The passages that support the answer are attached with document name and page.

  5. 05

    Verified response

    A reviewer confirms the answer, an approver signs it off, and the export carries the evidence.

EvidraDrafting job, 13 of 13
Drafting from the knowledge library13 / 13
  • 1.13 sourcesDrafted
  • 1.24 sourcesDrafted
  • 1.31 sourceDrafted
  • 2.12 sourcesDrafted
  • 2.23 sourcesDrafted
  • 2.32 sourcesDrafted
  • 2.43 sourcesDrafted
  • 2.52 sourcesDrafted
  • 3.12 sourcesDrafted
  • 3.23 sourcesDrafted
  • 4.12 sourcesDrafted
  • 4.2No evidence found
  • 4.34 sourcesDrafted
EvidraDue diligence questionnaire, question 1.2, drafted
1.2AIHigh confidence79%Drafted

How does the platform prevent AI generated answers that are not supported by the customer's documents?

Format: Free text
Answer

Our platform prevents AI-generated answers not supported by customer documents by ensuring that for each question, only passages retrieved from the organization's own document library are sent to the drafting model. Drafted answers are produced from these retrieved passages and are shown with the passages they came from. If no supporting passage is found, the question is marked "No evidence found." Additionally, drafted answers are a starting point for reviewers, and organizations are responsible for reviewing every answer before it is sent to a third party, with evidence attached for human checking.

Evidence4 sources
  • Evidra security overviewpage 2relevance 79%
    ... What each model receives: document text goes to the embedding model to build the organization's index, the text of a Word or PDF questionnaire goes to the drafting model to extract its questions, and for each question only the passages retrieved for it go to the drafting model. Subprocessors ...
  • Evidra security overviewpage 1relevance 71%
    ... organization, with public access prevented. - Vertex AI on Google Cloud, in the Netherlands region, runs the models: Google's text embedding model for passages and questions, and a Gemini model that drafts answers and reads the questions out of Word and PDF questionnaires. ...
  • Evidra terms of servicepage 1relevance 59%
    ... A question with no supporting passage is marked No evidence found. Drafts are a starting point for your reviewers: your organization is responsible for reviewing every answer before it is sent to a third party, and for the accuracy of what it sends. ...
  • Evidra acceptable use policypage 1relevance 54%
    ... Answers and evidence The evidence attached to an answer is there so a person can check it. Do not present an answer marked Unverified, or an answer edited beyond what its evidence supports, to a third party as a verified statement without reviewing it. ...
Module 4

Verification and export

Contributors, Reviewers and Approvers work on the same questionnaire with comments and an activity trail on every answer. Approved answers join the answer library for the next questionnaire.

EvidraVendor security assessment, question 2.3, approved
2.3Data protectionHigh confidence83%Approved

List the subprocessors that process customer data and the purpose of each.

Format: Free text
Answer

We use Google Cloud for authentication (Firebase Authentication), application database (Firestore), file storage (Cloud Storage), and for passage embeddings, question extraction, and answer drafting with Gemini models (Vertex AI). We also use Vercel for web hosting, serverless compute, and edge network services.

Evidence2 sources
  • Evidra security overviewpage 2relevance 82%
    ... - Google Cloud: Authentication (Firebase Authentication), application database (Firestore), file storage (Cloud Storage), and passage embeddings, question extraction and answer drafting with Gemini models (Vertex AI). Region: European Union (Firestore and Cloud Storage in the ...
  • Evidra privacy policypage 2relevance 64%
    ... Changes to this list are published on the security page before they take effect. Google Cloud: Authentication (Firebase Authentication), application database (Firestore), file storage (Cloud Storage), and passage embeddings, question extraction and answer drafting with Gemini models (Vertex AI). ...
Activity trail
  1. Evidra teamdrafted this answer from the knowledge library: 2 sources, high confidence
  2. Evidra teamapproved the answer
  3. Evidra teamsaved the approved answer to the answer library
EvidraExport, vendor-security-assessment.xlsx
Format
vendor-security-assessment-response.xlsx, answers written to column D of the Security assessment sheet, yes or no to column C
Questions
13 questions: 12 approved, 1 marked No evidence found and left blank
Evidence appendix
Appendix sheet with 31 cited passages, each with its document title and page
Activity
Export recorded in the activity trail with the approver's name and time

Export writes answers back into the original Excel workbook, or into a Word document for Word and PDF questionnaires, and adds an evidence appendix that lists every cited passage with its document title and page number. The approver's name and the export time are recorded in the activity trail.

How our AI works

An AI pipeline that has to show its sources

The same stages run on every questionnaire. The models read and write, but a rule in code decides what reaches your reviewers: no citation, no answer.

  1. 01

    Reads every page

    Policies, past responses and product documents are read page by page and split into passages that keep their page numbers, so every answer can point to the exact page it came from.

  2. 02

    Finds meaning, not just keywords

    Google text embeddings and keyword ranking work together in a hybrid search over one isolated index per organization. A question finds the right passage even when it uses different words.

  3. 03

    Reads the questionnaire for you

    Spreadsheets are parsed for sections, numbering, yes or no columns and character limits. Gemini reads Word and PDF questionnaires and returns each question with its section and answer format.

  4. 04

    Drafts only from your evidence

    Gemini writes each answer from the retrieved passages alone, in your organization's tone and inside the question's format and character limit, and cites the passages it used.

  5. 05

    Checks every citation in code

    A draft that cites nothing it retrieved is rejected before anyone sees it. The question is marked No evidence found, with a note on which document would answer it.

  6. 06

    Scores confidence and reuses approved answers

    Each answer carries a confidence signal from retrieval strength and citation coverage. Approved answers are matched to new questions by meaning and reused while their sources stay current.

The models are Gemini and Google text embeddings on Vertex AI, in a European Union region. Customer documents are never used to train them. How your data is handled

Start your response library

Upload the documents your team already trusts. Every answer that follows is drafted from them, cited and reviewed before it leaves.

Start your response libraryTalk to sales