Skip to main content

Evidence Management

Last Updated: March 27, 2026 (NQU-401 review)

Overview​

Evidence is the foundation of any inquiry. Nquiry provides comprehensive tools for uploading, organizing, and analyzing evidence items while maintaining the metadata needed for professional documentation.

Evidence Types​

Nquiry supports seven evidence types, each with appropriate handling:

TypeIconDescriptionBest For
Document📄PDFs, Word docs, spreadsheets, text filesPolicies, records, correspondence
Interview🎤Transcripts, recordings, statementsWitness accounts, depositions
Website🌐URLs, screenshots, web contentOnline sources, public records
Observation👁️Field notes, site visit documentationPhysical inspections, site visits
Dataset📊Structured data, exports, logsSystem data, statistical evidence
Note📝General notes, summaries, memosInternal documentation
Standard📏Policies, regulations, compliance criteriaNormative/compliance documents, IRAC pattern

Creating Evidence​

Required Fields​

  • Title: Descriptive name for the evidence
  • Type: One of the seven evidence types
  • Investigation: Auto-assigned to current project

Optional Fields​

  • Description: Summary of what the evidence contains
  • Content Text: Full text content or transcript
  • Source: Where the evidence came from
  • Source Date: When the evidence was created/obtained
  • Investigator Notes: Private notes about the evidence
  • To Be Collected: Flag for planned but not-yet-obtained evidence

File Attachments​

Each evidence item can have multiple file attachments:

  • PDFs, images, Word documents, Excel files
  • Files are stored securely in encrypted S3 storage
  • Text is extracted from PDFs and Office documents for search

Evidence Collection Tracking​

"To Be Collected" Feature​

Mark evidence items as "to be collected" when planning your investigation:

  • Create placeholder evidence items during planning
  • Flag them as "to be collected"
  • Add notes about how/when to obtain
  • System tracks collection progress

When evidence content is added (text or attachments), the "to be collected" flag automatically clears.

Collection Status Dashboard​

The Overview page shows:

  • Total evidence items
  • Items with attachments
  • Items pending collection
  • Evidence linked to questions

Linking Evidence to Questions​

Linking evidence to questions:

  • Ensures all questions have supporting evidence
  • Identifies evidence gaps
  • Provides context for AI analysis
  • Enables traceability in reports

From the evidence detail dialog:

  1. Click "Link to Questions"
  2. Select relevant questions
  3. Add optional relevance notes
  4. Save links

From the Questions page:

  1. Open a question
  2. View linked evidence
  3. Add new links

Evidence cards show:

  • Number of linked questions (badge)
  • Unlinked evidence highlighted in filters
  • Questions without evidence flagged in analysis

Evidence for AI Analysis​

How Evidence is Used​

When generating AI analysis:

  1. Semantic Search: System finds evidence most relevant to the question
  2. Context Building: Relevant chunks are assembled for AI review
  3. File Processing: PDFs and documents are read by the AI
  4. Citation Tracking: AI cites specific evidence in its analysis

Evidence is automatically processed for semantic search:

  • Text content is chunked into smaller segments
  • Chunks are converted to vector embeddings
  • Embeddings enable similarity-based retrieval

Processing status is shown in analysis generation feedback.

File Type Support​

The AI can directly analyze:

  • PDFs: Full document analysis with page references
  • Images: Visual analysis (JPEG, PNG, GIF, WebP)
  • Word Documents: Text extraction and analysis
  • Excel Files: Sheet-by-sheet text extraction

Evidence Organization​

Filtering​

Filter evidence by:

  • Type: Document, interview, website, etc.
  • Link Status: All, linked to questions, not linked
  • Collection Status: Collected, to be collected

Sorting​

Sort evidence by:

  • Newest first (default)
  • Oldest first
  • Source date
  • Type
  • Title (A-Z)

Bulk Upload​

Upload multiple evidence files at once via the bulk upload dialog. The dialog supports drag-and-drop and shows file details in a wider layout for comfortable review.

In-App File Viewer​

View evidence attachments and background documents directly in the app without downloading:

  • Supported formats: PDFs (rendered in browser PDF viewer), images (PNG, JPEG, GIF, WebP, SVG), text files (plain text, markdown, CSV, HTML)
  • Where it appears: "View" link next to "Download" on evidence detail pages; eye icon button on background documents in Settings
  • Unsupported formats: .docx, .xlsx, and other binary formats show only the Download option
  • Files are served securely through the app with 5-minute caching

Evidence Notes​

Investigator Notes​

Private notes attached directly to evidence:

  • Not included in reports by default
  • Useful for tracking follow-up items
  • Visible only to team members

Scoping AI Analysis with Evidence Notes (Evidence Windowing)​

Investigator notes serve a second purpose beyond annotation — they can be used to scope which portion of an evidence document gets sent to the AI during analysis. This is called evidence windowing.

The full document text is always preserved in the database and used for search and retrieval. Windowing only affects what gets loaded into the AI prompt at analysis time, keeping prompts focused and reducing noise.

Spreadsheet files (.xlsx, .xls, .csv)

Add range hints to the evidence notes field to scope which rows, columns, or sheets are sent to the AI:

HintExampleEffect
Cell rangeB1:D20Sends columns B–D, rows 1–20
Sheet + rangeSheet1!B1:D20Scopes to a specific sheet and range
Row rangerows 5-20Sends only rows 5 through 20 (header always included)
Column rangecolumns A-FSends only columns A through F
Sheet filterSheet2 onlySends only the named sheet

PDF files

Add page hints to scope which pages are sent to the AI:

HintExampleEffect
Page rangepages 12-15Sends pages 12 through 15
Single pagepage 5Sends only page 5
Abbreviated rangepp. 5-10Sends pages 5 through 10
Short formp. 7Sends only page 7

Example workflow — procedure compliance: Upload a multi-page procedure document as evidence. In the investigator notes, write page 5. When analysis runs, only page 5 is sent to the AI in the prompt. In the Direction for AI field on the Generate Analysis dialog, add the analytical instruction: "Determine whether the subject followed the sterilization procedure on page 5." The windowed content and the investigator direction work together for a focused, precise analysis.

Key behaviors:

  • Windowing hints are detected automatically — no special syntax beyond the natural language patterns above
  • Multiple hints are supported (e.g., Sheet2 only, rows 1-20)
  • If a requested range exceeds the document size, available content is returned without error
  • Spreadsheet hints on PDF files are ignored (and vice versa)
  • A note is prepended to the prompt when windowing is applied: "[Note: Content windowed per investigator guidance. Full data preserved for search.]"
  • Documents without windowing hints pass through unchanged

Implemented: NQU-298 (spreadsheets), NQU-314 (PDFs)

Evidence Notes (Annotations)​

Structured annotations on evidence:

  • Observation: What the evidence shows
  • Analysis: Interpretation of significance
  • Follow-up: Additional work needed
  • Concern: Flags or red flags
  • General: Other notes

Security and Compliance​

Storage Security​

  • Files encrypted at rest (AES-256)
  • Files encrypted in transit (TLS 1.2+)
  • Stored in organization-isolated S3 paths
  • Access requires authentication + org authorization

Audit Logging​

All evidence actions are logged:

  • Upload, view, download, delete
  • Link/unlink to questions
  • User, timestamp, IP address
  • Success/failure status

HIPAA Considerations​

  • PHI detection available via Bedrock Guardrails
  • Evidence access restricted to org members
  • Audit trail maintained for compliance

Evidence Evaluation​

The 10 CIGIE/GAO quality criteria below shape what the AI looks for when assessing evidence. They're used as a conceptual framework, not as individually-scored dimensions — per NQU-100, the active output is a simpler 2-field assessment (confidence + reasoning) rolled up across all 10.

  1. Relevance - Logical connection to the question
  2. Reliability - Trustworthiness of source
  3. Sufficiency - Adequate quantity and scope
  4. Validity - Accurately represents what it claims
  5. Competence - Quality appropriate to form
  6. Completeness - No critical gaps
  7. Timeliness - Current and applicable
  8. Objectivity - Fact-based vs. opinion
  9. Authenticity - Genuine and verifiable
  10. Consistency - Aligns with other evidence

See Evidence Evaluation Framework for the historical reference.

Best Practices​

Collection​

  • Document source and date for every item
  • Use consistent naming conventions
  • Attach original files when possible
  • Record chain of custody for physical evidence

Organization​

  • Link evidence to questions as you collect
  • Use investigator notes for follow-up items
  • Filter regularly to find unlinked items
  • Review "to be collected" items weekly

Quality​

  • Verify authenticity before relying on evidence
  • Seek corroboration for critical claims
  • Note any concerns about reliability
  • Document chain of custody

Three-Stage Hybrid Retrieval Pipeline​

Added 2026-03-06 (NQU-379). Major update 2026-03-27 (NQU-462).

Nquiry uses a three-stage hybrid retrieval approach:

Stage 1 — Keyword Search. PostgreSQL tsvector/tsquery finds evidence containing the same terms as the question. Catches exact-match evidence (case numbers, names, policy codes, identifiers) that semantic search misses. Canonical example: NQU-380 DOC-198 found exclusively by keyword search.

Stage 2 — Semantic Search (Vector Embeddings). Amazon Titan Text Embeddings V2 via pgvector converts questions and evidence into meaning vectors. Catches evidence that's relevant even when it uses different words.

Stage 3 — Reranking. Cohere Rerank 3.5 re-evaluates merged results from Stages 1 and 2 by reading question and evidence together. Promotes keyword-surfaced results that semantic search missed; demotes false positives.

Team-draft interleaving (NQU-462): Results from keyword and vector search are combined using alternating picks (team-draft), replacing the old ad-hoc "guaranteed keyword top-20" slot reservation. This scales naturally without hardcoded slot counts.

Query-adaptive weighting (NQU-462): The system automatically classifies queries as entity-heavy (proper nouns, dates, IDs → 2.5x keyword boost), conceptual (abstract → 1.0x), or hybrid (moderate → 1.5x). No LLM needed — uses heuristic detection of capitalized multi-word sequences and specific identifiers.

Continuous feedback (NQU-462): After each analysis generation, the system computes retrieval quality by comparing retrieved evidence vs. AI citations, storing results in the retrieval_benchmark_result table for trend monitoring.

Baseline metrics (Davenheim, 5 queries): R@5=0.166, R@10=0.283, R@20=0.323, MRR=0.400, NDCG@20=0.362.

Storage and Quotas​

PlanStorage Limit
Trial2 GB
Professional50 GB
Team200 GB

Storage used is shown in Settings → Billing.

Large files (>25MB) may require splitting or compression.