Understanding How Nquiry Works
A Guide for Oversight, Compliance, and Inquiry Professionals
Document Purpose: This guide provides professionals with a clear understanding of how Nquiry handles evidence, generates AI-assisted analysis, and supports their work. The goal is to give you the information you need to use Nquiry with confidence.
Who This Is For: Investigators, auditors, inspectors, compliance officers, quality reviewers, and anyone conducting formal inquiries who wants to understand what's happening "under the hood" when they use AI-assisted analysis tools.
Table of Contents
- The Problem Nquiry Solves
- How Your Data Is Organized and Stored
- Data Security and Isolation
- How Evidence Becomes Searchable
- How AI Analysis Works
- The Evidence Evaluation Framework
- Investigator Direction and Control
- How Analysis Builds on Analysis
- How Nquiry Checks Its Own Work
- Understanding Quality Metrics in Detail
- Evidence Readiness and Coverage
- Your Role: The Human in the Loop
- What Nquiry Cannot Do
- Frequently Asked Questions
- Glossary
1. The Problem Nquiry Solves
The Reality of Modern Inquiries
Professionals who conduct inquiries — whether federal oversight reviews, healthcare quality assessments, corporate compliance investigations, or internal audits — face a fundamental challenge: the volume of evidence has grown exponentially, but the time and resources available to analyze it have not.
Consider a typical inquiry: you might have hundreds of emails, financial records spanning years, interview transcripts, policy documents, and supporting documentation. Each piece of evidence needs to be reviewed, its relevance assessed, and its relationship to your inquiry questions understood.
Traditional approaches require professionals to hold this entire body of evidence in their minds while methodically working through analysis. This is cognitively demanding, time-consuming, and prone to human limitations around memory, attention, and consistency.
What AI Assistance Offers
Nquiry uses AI to help with the analysis phase of inquiries — not to replace professional judgment, but to augment it. Specifically, the AI can process large volumes of text quickly, find relevant evidence across your entire evidence collection, apply consistent evaluation criteria to each piece of evidence, identify patterns and connections you might miss, generate draft analyses that you review and refine, and flag gaps in your evidence coverage.
The key word is assist. The AI generates analysis based on the evidence you've collected and the questions you've defined. You review that analysis, verify it against your evidence, and make the final determinations.
What Makes Nquiry Different from Generic AI
You may have encountered AI-generated content that feels hollow — generic summaries that could apply to any topic, confident assertions without substance, or responses that sound authoritative but say nothing.
Nquiry is designed to avoid this through specific architectural choices. Every claim in the analysis must be traceable to specific evidence you provided. The AI only uses evidence from your inquiry — not external knowledge or training data. Confidence levels are explicitly stated and evidence gaps are identified. The system documents which evidence supports each conclusion. A structured evidence evaluation framework grounded in professional oversight standards governs all analysis. Automated quality checks independently verify the AI's work. And you can direct the AI's analysis focus using your professional judgment.
The rest of this document explains exactly how these safeguards work.
2. How Your Data Is Organized and Stored
Understanding how Nquiry organizes your data helps you understand what the AI has access to when generating analysis.
The Inquiry Structure
Every inquiry in Nquiry follows a hierarchical structure:
Inquiry (your case, review, audit, or project)
├── Focus Statement (what you're examining)
├── Topics (logical groupings of issues)
│ └── Questions (specific things to determine)
│ └── Investigator Direction (your guidance for the AI)
├── Evidence
│ ├── Documents (PDFs, emails, records)
│ ├── Interviews (transcripts, statements)
│ ├── Observations (field notes, site visits)
│ ├── Datasets (structured data, logs)
│ ├── Websites (URLs, screenshots)
│ └── Notes (your documentation)
├── Background Documents (context: charge letters, org charts, scope memos)
├── Framework Documents (evaluation criteria: policies, regulations, standards)
└── Analyses
├── Question Analyses (per-question evidence evaluation)
├── Topic Analyses (cross-question synthesis)
├── Overall Summary (investigation-level synthesis)
├── Gap Analysis (systematic evidence gap identification)
└── Error Check (cross-analysis consistency verification)
Work Types
Nquiry supports multiple types of professional work, including evaluations, reviews, audits, inspections, investigations, inquiries, assessments, and cases. The structure adapts to your specific needs.
Evidence Linking
A critical feature of Nquiry is the ability to link evidence items to specific questions. This serves two purposes: traceability (you can see which evidence supports or contradicts each question) and AI context (the AI prioritizes linked evidence when analyzing a specific question).
When you link evidence to a question, you're telling the system "this evidence is relevant to answering this question." The AI uses this information alongside its own semantic search to find relevant evidence. If you've linked evidence that the AI's search wouldn't have found on its own, the link divergence indicator lets you know — this is a feature, not a problem, since your professional judgment may identify connections the AI can't.
3. Data Security and Isolation
Organization Boundaries
Your data is isolated within your organization. This isolation is enforced at the database level — queries are automatically filtered to only return data belonging to your organization. A user from Organization A cannot access, search, or view any data from Organization B, even if they somehow obtained a direct database identifier.
Access Control
Within your organization, access is controlled by roles:
| Role | Can Create Inquiries | Can View All Inquiries | Can Edit All Inquiries | Can Manage Members | Can Access Billing |
|---|---|---|---|---|---|
| Owner | ✓ | ✓ | ✓ | ✓ | ✓ |
| Admin | ✓ | ✓ | ✓ | ✓ | |
| Member | ✓ | ✓ | Own only | ||
| Viewer | ✓ |
Audit Trail
Every significant action in Nquiry is logged: who accessed what evidence, when analyses were generated, what data was exported, authentication events, and changes to inquiry structure. These logs support compliance requirements and enable you to demonstrate exactly who accessed case data and when.
Infrastructure
Nquiry runs on AWS cloud infrastructure in the United States. All data is encrypted at rest (AES-256) and in transit (TLS 1.2+). Authentication is managed through a dedicated identity service with support for multi-factor authentication.
Nquiry is designed to support HIPAA compliance for healthcare-related inquiries, with appropriate technical safeguards and Business Associate Agreement coverage available.
4. How Evidence Becomes Searchable
This section explains the process that makes your evidence searchable by the AI.
Text Extraction
When you upload evidence, Nquiry extracts text content from supported file types including PDFs (with OCR for scanned documents), Word documents, Excel spreadsheets, images (via AI visual analysis), and plain text files.
Chunking: Breaking Text Into Pieces
AI models have limits on how much text they can process at once. Additionally, retrieving small, focused passages is more effective than sending entire documents.
Nquiry splits your evidence text into segments called "chunks." These chunks overlap slightly so important context at boundary points isn't lost, respect natural boundaries like paragraphs and sentences when possible, and retain source information so any chunk can be traced back to its original document and location within that document.
Making Text Searchable by Meaning
Traditional keyword search finds documents containing specific words. But inquiries often need to find relevant content even when the exact words differ.
For example, you might ask: "Did the employee access records outside business hours?" Relevant evidence might include an email saying "logged in at 11 PM," a system log showing "authentication: 23:14:22," or a statement mentioning "working late on personal projects." None of these contain the words "outside business hours," but all are semantically relevant.
Nquiry uses semantic search. Each chunk of evidence is converted into a mathematical representation (called an "embedding") that captures its meaning. When analyzing a question, that question is also converted into an embedding, and the system finds evidence whose meaning is mathematically similar to the question.
This embedding process happens within Nquiry's secure AWS infrastructure using Amazon Titan Embeddings V2. Your evidence is not sent to external services.
How AI Summaries Help You Navigate Evidence
When evidence is uploaded and processed, Nquiry generates a brief AI summary of each evidence item. This summary appears on evidence cards in the interface, giving you a quick sense of what each item contains without opening it. The summary also notes how many questions the evidence is linked to, helping you spot evidence that may be underutilized.
5. How AI Analysis Works
This section walks through exactly what happens when you request an AI analysis.
Before Analysis: Evidence Readiness Check
Before running the AI, Nquiry assesses whether your evidence is ready for analysis. This happens in seconds and shows you a readiness assessment that includes a readiness score (0-100), the count of relevant evidence items identified, a source breakdown showing what types of evidence the AI will consider, any coverage gaps that might limit the analysis, and recommendations for improving readiness (such as uploading additional evidence or linking evidence to questions).
This check does not make API calls to the AI — it's a quick analysis of your evidence collection's completeness. You can cancel here if the readiness assessment suggests you need more evidence before proceeding.
What the AI Receives
When you proceed with analysis for a question, the AI receives:
- Your question — the specific question to be answered
- Your direction (if provided) — any guidance you've set (see Section 7)
- Inquiry context — title, focus statement, type of work, investigation phase
- Retrieved evidence — the most relevant evidence chunks found through semantic search
- Linked evidence — any evidence you've manually linked to this question
- Background documents — context documents you've marked "include in analysis" (charge letters, org charts, scope memos)
- Framework documents — policies, standards, or criteria for evaluation (organizational policies, regulations, professional standards)
- The Evidence Evaluation Framework — Nquiry's built-in structured instruction set (see Section 6)
- Prior analysis context — for topic-level and summary analyses, the AI also receives completed question-level analyses to build upon (see Section 8)
Context Budget: How Evidence Is Prioritized
The AI can only process a limited amount of text at once (the "context window"). For investigations with large amounts of evidence, not everything can fit. Nquiry manages this through a context budget allocation:
Evidence is ranked by relevance (semantic similarity to the question + manual links). The most relevant evidence is included first. Background documents and framework documents receive dedicated portions of the budget. When the budget is full, remaining evidence is excluded but logged — you can see exactly what was and wasn't included in the Evidence Considered panel.
For most investigations (fewer than 50 evidence items), all relevant evidence fits comfortably. For larger investigations, the quality metrics (Section 10) help you understand whether important evidence was excluded.
What the AI Does Not Receive
The AI does not receive evidence from other inquiries, evidence from other organizations, data from the internet or external sources, your personal information unrelated to the inquiry, or previous conversations or chat history. Each analysis request starts fresh with only the context explicitly provided.
The Analysis Process
The AI follows a structured process:
Step 1: Identify Question Elements. Break down your question into specific factual elements that must be established. For example, "Did Manager X retaliate by giving Employee Y a negative review?" breaks into: Did Employee Y engage in protected activity? Did Manager X give a negative review? Was there a temporal and causal connection between the two? Was Manager X's stated reason pretextual?
Step 2: Inventory Evidence. List and categorize each piece of evidence by type (documentary, testimonial, digital, physical) and assess its relationship to each question element.
Step 3: Evaluate Each Piece. For each piece of evidence, assess it using the evidence evaluation framework (see Section 6) and assign a quality rating:
| Rating | Meaning |
|---|---|
| Strong | Primary source, corroborated, contemporaneous, objective, verifiable |
| Moderate | Reliable source with some limitations (minor corroboration gaps, some subjectivity) |
| Weak | Significant limitations: hearsay, uncorroborated, retrospective, or potentially biased |
| Insufficient | Cannot be evaluated or does not meaningfully contribute to analysis |
Step 4: Synthesize. For each question element, identify which evidence addresses it, assess whether multiple sources converge or conflict, identify any contradictions, and determine whether the evidence is sufficient to reach a conclusion.
Step 5: State Conclusion with Confidence Level.
| Confidence Level | Meaning |
|---|---|
| Established | Multiple strong, convergent evidence items. Alternative explanations addressed. Quality checks verified. |
| Probable | Good evidence with minor gaps. Faithfulness ≥ 85%, coverage ≥ 85%. Reasonable basis for conclusion. |
| Possible | Some supporting evidence but significant gaps, weaknesses, or unresolved contradictions. |
| Insufficient | Evidence too weak, sparse, or conflicting to support a conclusion. |
Step 6: Document. The AI outputs structured analysis showing the conclusion and confidence level, how each piece of evidence was evaluated (with quality ratings), what gaps exist, what alternative explanations were considered, and what the AI recommends for follow-up.
What You See
The generated analysis appears with: the conclusion text, a confidence level indicator, a quality badge summarizing automated quality checks, citations to specific evidence with evaluation details, identified gaps and issues, the trust layer — a strip showing context coverage, framework application status, and link divergence at a glance, and an expandable Evidence Considered panel showing exactly which evidence the AI reviewed, including similarity scores and included/excluded status.
6. The Evidence Evaluation Framework
At the heart of Nquiry's AI analysis is a formal evidence evaluation framework. This isn't generic AI prompting — it's a comprehensive 600+ line instruction set based on cross-sector professional standards from CIGIE (Council of the Inspectors General on Integrity and Efficiency), GAO (Government Accountability Office), IIA (Institute of Internal Auditors), and ACFE (Association of Certified Fraud Examiners).
How Evidence Is Evaluated
For each piece of evidence, the AI produces a structured assessment with two fields:
- Relevant (yes/no) — Does this evidence address the question being asked?
- Has Limitations (yes/no) — Does this evidence have material limitations that affect its weight?
Each assessment includes a free-text rationale explaining the AI's reasoning. This simplified model replaced an earlier 10-criteria checklist after validation testing showed it produced substantially better inter-rater agreement (κ improved from -0.008 to 0.720).
Quality Dimensions That Inform the Evaluation
The AI's assessment draws on ten quality dimensions from professional oversight standards. These dimensions are not individually scored — they inform the AI's judgment about relevance and limitations:
Relevance — Does this evidence have a logical connection to the question being examined? The AI considers whether the evidence addresses the matter, falls within the correct time period, and involves the relevant parties.
Reliability — Is the source trustworthy and the evidence free from bias or error? The AI assesses source independence, corroboration by other sources, whether it was created at the time of events, and whether it's an original document versus hearsay.
Sufficiency — Is there enough evidence to support a conclusion? Multiple sources agreeing is stronger than a single source on a critical point. The AI identifies when coverage is adequate versus when gaps prevent meaningful conclusions.
Validity — Does this evidence actually prove what it's offered to prove? Direct evidence is weighted more heavily than circumstantial evidence that requires inference.
Competence — Is the source qualified? Was the evidence properly created or collected? The AI considers expert credentials, established business processes, and proper handling procedures.
Completeness — Does the evidence cover the matter thoroughly? The AI looks for missing categories of evidence, gaps in time periods, and whether both supporting and contradictory evidence is present.
Timeliness — Was the evidence created or obtained at an appropriate time? Contemporary records carry more weight than retrospective accounts.
Objectivity — Is this fact-based or opinion-based? Observable facts and verifiable data are weighted more heavily than speculation or subjective characterizations.
Authenticity — Is this evidence genuine and verifiable? The AI considers whether origin can be verified, whether chain of custody is documented, and whether there are signs of alteration.
Consistency — Does this align with other evidence? Multiple independent sources reporting the same facts strengthens confidence. Unexplained contradictions reduce it.
Evidence Type Considerations
The framework provides specific guidance for evaluating different evidence types:
Testimonial evidence (interviews, statements) is evaluated for whether the witness has direct knowledge versus reporting what others said (hearsay), indicators of credibility or bias, consistency with other witnesses' accounts, and whether the statement was made close in time to the events described.
Documentary evidence (emails, records, memos) is assessed for whether records were created in the normal course of business, whether they're contemporaneous to events, whether they're originals versus copies or summaries, and chain of custody.
Digital evidence (logs, metadata, system records) requires integrity verification, proper preservation, and consideration of whether systems could have been manipulated.
Expert evidence (analysis, opinions, assessments) is evaluated for relevant qualifications, whether the methodology is sound and generally accepted, and whether the conclusions are supported by the underlying data.
How Framework Documents Enhance Analysis
When you upload framework documents — organizational policies, regulations, professional standards, criteria — the AI integrates them into its analysis. For example, if you're investigating whether a healthcare provider followed prescribing guidelines, uploading those guidelines as a framework document lets the AI evaluate the evidence specifically against each guideline requirement.
The AI applies a relevance check before using framework documents: only criteria that are directly applicable to the question and evidence are applied. If a framework document isn't relevant to a particular analysis (for instance, an HR policy in a financial audit question), the AI explicitly states that the framework was reviewed but not applied, and explains why.
7. Investigator Direction and Control
One of Nquiry's most important features is letting you direct the AI's analysis rather than accepting a one-size-fits-all approach.
Setting Direction on Questions
Each question in your inquiry has a "direction" field where you can provide guidance to the AI. This direction is included in the AI's context when generating analysis for that question.
Effective direction examples:
- "Focus on timeline discrepancies between the badge swipe data and the encounter logs"
- "Pay particular attention to the credibility of Witness A given their relationship to the subject"
- "Evaluate whether the documentation practices meet the requirements in Section 4.2 of the policy"
- "The key issue is whether the pattern of prescribing changed after the peer review, not whether individual prescriptions were improper"
The AI treats your direction as a priority instruction that shapes — but does not replace — its systematic evidence evaluation. It will still evaluate all relevant evidence, but will weight its analysis toward the areas you've identified as most important.
Direction Categories
Nquiry recognizes several types of investigator direction:
Focus narrowing — directing the AI to concentrate on specific aspects of a question ("focus on the financial records, not the interview testimony").
Evidence weighting — indicating that certain evidence should receive more attention ("the PDMP data is the most probative evidence for this question").
Analytical approach — specifying how you want the analysis structured ("compare the subject's statements chronologically against the documentary evidence to identify inconsistencies").
Alternative explanations — directing the AI to consider specific alternatives ("evaluate whether the pattern could be explained by the change in patient panel rather than intentional overbilling").
What Direction Cannot Do
Direction cannot instruct the AI to ignore evidence, reach a predetermined conclusion, omit gaps or limitations from its analysis, or rate evidence quality higher or lower than the framework criteria support. The AI will note if a direction conflicts with its evidence evaluation.
8. How Analysis Builds on Analysis
Nquiry generates different types of analysis that build on each other in a deliberate sequence.
The Analysis Hierarchy
Question Analyses (foundation — one per question)
↓ feeds into
Topic Analyses (synthesis — one per topic, draws on its question analyses)
↓ feeds into
Overall Summary (investigation-level synthesis — draws on all topic analyses)
Gap Analysis (cross-cutting — identifies evidence gaps across all questions)
Error Check (quality assurance — identifies inconsistencies across all analyses)
Why This Matters
When the AI generates a topic analysis, it receives the completed question-level analyses for that topic's questions as input. This means the topic analysis doesn't re-analyze all the raw evidence from scratch — it synthesizes the findings, identifies patterns across questions, flags contradictions between question-level conclusions, and identifies gaps that only become visible when viewing findings together.
The overall summary then synthesizes topic-level findings into investigation-wide conclusions.
This chaining approach has several advantages. Each analysis level adds insight rather than repeating work. Contradictions between question-level analyses are explicitly surfaced at the topic level. The overall summary reflects the full analytical path, not just a single pass through all evidence. And each level maintains its own quality metrics, so you can identify exactly where in the chain a quality issue originates.
The Analysis Record
Every analysis in Nquiry records which prior analyses it built upon, what evidence was in its context, what prompt version was used, and its full quality metrics. This creates a complete analytical provenance chain.
9. How Nquiry Checks Its Own Work
Generating AI analysis is only half the story. After every analysis, Nquiry automatically runs independent quality checks.
Faithfulness Check
What it does: A separate AI process reads the generated analysis, identifies each factual claim, and checks whether the evidence in context actually supports that claim.
Why it matters: This catches "hallucination" — when the AI makes claims that sound plausible but aren't supported by your evidence. Every unsupported claim is flagged for your review.
What you see: A faithfulness score (0-100%) and, if any claims are unsupported, a list of specific claims to verify.
Coverage Check
What it does: A separate AI process breaks your question into its component elements and checks whether the analysis addresses each one.
Why it matters: This catches incomplete analysis — when the AI addresses part of your question thoroughly but misses other parts entirely.
What you see: A coverage score (0-100%) and a list of any question elements not adequately addressed.
Schema Validation
What it does: Verifies that the AI output conforms to Nquiry's required structure — all required fields present, data types correct, quality ratings assigned to each evidence item.
Why it matters: Ensures the analysis can be properly displayed, stored, and used in reports. If validation fails, you see a clear error rather than malformed output.
Quality Confidence Badge
The results of all checks are summarized in a quality confidence badge:
| Badge | Meaning |
|---|---|
| Established (green) | High faithfulness + high coverage + validation passed. Reliable analysis. |
| Probable (blue) | Good scores with minor gaps. Usable with normal review. |
| Possible (amber) | Notable limitations. Review carefully before relying on conclusions. |
| Insufficient (gray) | Significant quality concerns. Consider regenerating with more evidence or direction. |
10. Understanding Quality Metrics in Detail
This section explains each quality metric in depth — what it measures, how it's calculated, and what the numbers mean for your work.
Faithfulness Score
The question it answers: "Is the AI only making claims that the evidence supports?"
How it's calculated: After the analysis is generated, a second AI call reviews the output. It identifies every factual claim (e.g., "Dr. Chen's prescribing rate was 3x the departmental average" or "the complaint was filed on December 15"). For each claim, it checks whether the evidence chunks that were in the AI's context contain information supporting that claim. The score is the percentage of claims that are supported.
Score interpretation:
| Range | Meaning | What to do |
|---|---|---|
| 95-100% | Excellent. Virtually every claim traces to evidence. | High confidence in factual accuracy. Normal review. |
| 85-94% | Good. Minor unsupported claims, likely reasonable inferences. | Check the flagged claims — they may be reasonable inferences the check was conservative about. |
| 70-84% | Concerning. Some claims may not be directly evidence-based. | Review the analysis against source evidence. The AI may be drawing on background knowledge rather than your evidence. |
| Below 70% | Problem. Many claims lack evidence support. | Don't rely on this analysis without thorough verification. Consider regenerating with more specific direction. |
Why it might be low even when the analysis seems correct: The faithfulness check only verifies claims against the evidence that was in the AI's context at analysis time. If a claim is true but the supporting evidence wasn't retrieved (because it had low semantic similarity to the question), the check will flag it as unsupported. This is the retrieval quality metric's domain — see below.
Coverage Score
The question it answers: "Did the analysis address all parts of my question?"
How it's calculated: The question is decomposed into its component elements (what specific facts need to be established). The analysis text is then evaluated for whether each element is addressed. The score is the percentage of elements covered.
Score interpretation:
| Range | Meaning | What to do |
|---|---|---|
| 90-100% | Comprehensive. All aspects of your question were addressed. | The analysis is thorough. |
| 80-89% | Good. Minor elements may not be explicitly addressed. | Review the gaps list — they may be intentionally omitted because evidence was lacking, which is actually honest. |
| 65-79% | Partial. Notable gaps in what the analysis covers. | Check why elements were missed. Provide direction to address gaps, or split into multiple questions. |
| Below 65% | Incomplete. Major aspects of your question weren't analyzed. | The question may be too broad, or critical evidence may be missing. |
Why it might be low even with strong evidence: Complex questions with multiple elements are harder to fully cover. A question like "Did the subject falsify records, and if so, what was the financial impact, and did management have knowledge?" contains at least three distinct analytical elements. The AI may thoroughly address two and barely touch the third. Direction helps: "Ensure your analysis addresses management's knowledge specifically."
Context Coverage
The question it answers: "How much of my evidence did the AI actually see?"
What it shows: "Analyzed X of Y evidence items (Z%)"
This is different from the coverage score. Context coverage is a raw count: of all the evidence items in your investigation, how many were included in the AI's context window for this analysis?
Interpretation:
- 100% = The AI had access to all your evidence. For investigations with fewer than ~50 items, this is typical.
- 70-99% = Some evidence was excluded due to context budget limits. Check the Evidence Considered panel to see what was excluded and whether it matters.
- Below 70% = A significant portion of your evidence wasn't in context. This is expected for very large investigations and is handled by the analysis chaining approach (Section 8).
Retrieval Quality
The question it answers: "How relevant was the evidence the system found?"
How it's calculated: When the AI searches your evidence collection, each retrieved chunk receives a similarity score (0-1) indicating how closely its meaning matches your question. These scores are aggregated.
| Level | Average Similarity | Meaning |
|---|---|---|
| Strong | > 0.85 | The AI found highly relevant evidence. Your evidence collection directly addresses this question. |
| Moderate | 0.70 - 0.85 | Reasonably relevant evidence found. Some retrieved items may be tangential. |
| Weak | < 0.70 | The evidence collection may not directly address this question. |
What the detailed retrieval statistics tell you:
The Evidence Considered panel shows each retrieved chunk with its similarity score, color-coded:
- Green (≥ 0.85) = highly relevant
- Yellow (≥ 0.70) = relevant
- Orange (≥ 0.60) = marginally relevant
- Gray (< 0.60) = low relevance
You can also see: total chunks retrieved, how many were included vs excluded from the AI's context, and source breakdown by evidence type (documents, interviews, notes, etc.).
Link Divergence
The question it answers: "Did the AI find the same evidence relevant as I did?"
When you manually link evidence to a question and the AI's semantic search identifies different evidence as most relevant, that's a divergence. The link divergence indicator shows you when this happens.
This is informational, not a problem. You may have linked evidence because you know something about the investigation that a semantic search can't capture — a seemingly unrelated document that contains a critical admission, for example. The divergence indicator simply flags the difference so you can provide direction if needed: "The manually linked evidence (DOC-015) is relevant because it establishes Manager X's awareness of the policy, even though it's about a different incident."
Evidence Readiness Score
The question it answers: "Is my evidence collection ready for analysis on this question?"
How it's calculated: Before analysis generation, Nquiry evaluates your evidence against the specific question:
- How many evidence items are linked to this question?
- How many evidence items did semantic search identify as relevant?
- What types of evidence are represented (documents only? interviews? multiple source types)?
- Are there obvious gaps in evidence categories?
Score interpretation:
| Range | Meaning |
|---|---|
| 80-100 | Strong evidence base. Multiple source types, good coverage. |
| 60-79 | Adequate. Analysis will likely produce useful results, but some gaps may limit confidence. |
| 40-59 | Limited. Analysis will proceed but expect lower confidence and more identified gaps. |
| Below 40 | Insufficient. Consider collecting more evidence before generating analysis. |
You see this score in the readiness check before analysis begins, giving you the opportunity to pause and gather more evidence if needed.
11. Evidence Readiness and Coverage
This section explains the indicators that help you understand whether your evidence collection is adequate for the analysis you're requesting.
Coverage Gaps
When Nquiry identifies that certain types of evidence are missing for a question, it flags these as coverage gaps. Common gap types include:
Source type gaps — "All evidence for this question is documentary. No testimonial evidence (interviews, statements) is available." This matters because different evidence types provide different perspectives and corroboration.
Temporal gaps — "Evidence covers January through March but the question period extends through June." Missing time periods may mean important events are undocumented.
Perspective gaps — "Evidence includes the subject's statements but no independent corroboration from other witnesses." Single-source findings are inherently weaker.
Contradictory evidence — "Evidence on this point is conflicting, with documentary records and testimonial evidence pointing in different directions." The AI will note this, but resolving contradictions requires professional judgment about credibility and weight.
Source Diversity
The evidence readiness check evaluates source diversity — whether your evidence comes from multiple independent sources and types. Analysis based on a single document is inherently limited compared to analysis drawing on documents, interviews, digital records, and third-party verification.
How These Indicators Help Your Work
These indicators don't tell you your evidence is "wrong" — they tell you where your evidence is strongest and where additional collection might strengthen your conclusions. Think of them as a systematic version of the mental checklist experienced investigators use naturally: "Do I have enough? Have I heard from all sides? Am I missing a time period?"
12. Your Role: The Human in the Loop
You Are Not Replaced
Nquiry is designed around a principle: AI assists, humans decide. This isn't just a disclaimer — it's built into how the system works.
What the AI does: Processes evidence quickly, applies consistent evaluation criteria, generates draft analysis, identifies gaps and flags issues, cites evidence for each claim, and checks its own work against quality standards.
What you do: Define scope and questions, collect and organize evidence, provide direction for analysis, review AI analysis against source evidence, verify claims and assess credibility, make finding determinations, write final conclusions, and take responsibility for the work product.
The Verification Workflow
A responsible workflow with AI-assisted analysis:
- Structure your inquiry — Define questions that are specific and answerable
- Collect evidence thoroughly — The AI can only work with what you provide
- Link key evidence — Help the system understand what's relevant
- Set direction — Tell the AI what to focus on for each question
- Check readiness — Review the evidence readiness assessment before proceeding
- Generate analysis — Let the AI create a draft
- Check quality metrics — Review faithfulness, coverage, retrieval quality
- Verify citations — Confirm that cited evidence actually supports the claims
- Investigate gaps — Address identified evidence gaps
- Apply judgment — Make your own assessment of the evidence
- Accept, reject, or refine — Mark the analysis as accepted if it meets your standards, reject it if it doesn't, or provide refined direction and regenerate
- Assign finding status — You determine whether findings are substantiated
- Generate report — Build your report from accepted analyses, editing as needed
Finding Status: Your Determination
After reviewing analysis, you assign a finding status:
| Status | Meaning |
|---|---|
| Pending | Not yet reviewed or determined |
| Substantiated | You have determined the evidence clearly supports the finding |
| Not Substantiated | You have determined the evidence does not support the allegation |
| Inconclusive | You have determined the evidence is insufficient to conclude either way |
These statuses are yours to assign. The AI may suggest confidence levels, but the determination is your professional judgment.
When to Trust vs. Verify
Higher trust (still verify): High quality badge (Established), strong retrieval quality, multiple evidence sources corroborating, and claims that align with your understanding of the evidence.
Requires close verification: Lower quality badges, complex or sensitive matters, claims that would be surprising if true, contradictory evidence the AI may not have weighted correctly, and high-stakes situations.
Never skip verification for: Final conclusions in formal reports, matters with legal implications, personnel actions, public-facing findings, and cases where you have context the AI doesn't.
13. What Nquiry Cannot Do
Understanding limitations is as important as understanding capabilities.
Does not make determinations. Nquiry provides analysis; it does not make findings. You determine what findings are substantiated.
Does not access external information. The AI works only with evidence you've uploaded. It does not search the internet, access external databases, or know current events.
Does not guarantee accuracy. AI can make mistakes. Quality checks catch many issues but not all. The AI may misinterpret nuanced evidence, miss complex relationships, overlook visual elements, or lack organizational context you take for granted.
Does not process all file types. Currently supported: PDF, Word, Excel, common image formats, and plain text. Not currently supported: video files, audio files (transcribe externally and upload text), and specialized proprietary formats.
Does not provide legal advice. The AI applies evidence evaluation standards but is not a lawyer. For matters with legal implications, consult qualified legal counsel.
Does not replace professional skills. Some capabilities require human judgment that AI cannot replicate: assessing witness demeanor and credibility in interviews, conducting physical inspections, making judgment calls about inquiry direction, handling sensitive interpersonal situations, and exercising discretion.
Context window limitations. For very large evidence sets, not all evidence fits in a single analysis. The quality metrics indicate when this is occurring, and the analysis chaining approach (Section 8) addresses it by building analysis in layers.
14. Frequently Asked Questions
About AI and Trust
Q: Is the AI making up information that isn't in my evidence? The AI is instructed to only use evidence you've provided. The faithfulness check independently verifies this. Review the faithfulness score and any flagged unsupported claims.
Q: How do I know the AI understood my evidence correctly? Check the analysis against your original evidence. Citations reference specific evidence — verify they make sense. If something seems off, provide more specific direction and regenerate, or add clarifying evidence.
Q: What if the AI analysis contradicts my assessment? Your judgment takes precedence. The AI provides one perspective based on systematic evaluation. If you disagree, document your reasoning. You may have context the AI lacks.
Q: Can I use AI-generated analysis in official reports? Yes, with important caveats: review and verify the analysis, edit to reflect your conclusions, document that AI assistance was used per your organization's policies, and take responsibility for the final work product.
Q: What does it mean when the quality badge says "Possible" or "Insufficient"? These indicate the automated quality checks found limitations. "Possible" means notable gaps that warrant careful review. "Insufficient" means significant concerns — don't rely on the analysis without thorough verification. Check the specific faithfulness and coverage scores for details on what fell short.
Q: The Evidence Considered panel shows evidence was "excluded." Is that a problem? Not necessarily. When the context budget is full, lower-relevance evidence is excluded. Check the excluded items — if they're tangential, no concern. If important evidence was excluded, you may want to link it to the question directly (which forces inclusion) and regenerate.
About Data and Security
Q: Who can see my inquiry data? Only members of your organization with appropriate roles. Data is isolated at the database level.
Q: Is my data used to train AI models? No. Your inquiry data is processed to generate your analysis but is not used to train or improve AI models.
Q: How long is my data retained? Your data is retained as long as your account is active and you haven't deleted it. Audit logs are retained per applicable compliance requirements.
Q: Is Nquiry suitable for healthcare-related inquiries? Nquiry is designed to support HIPAA compliance with appropriate technical safeguards and Business Associate Agreement coverage.
About Using Nquiry Effectively
Q: How do I get the best analysis results? Write clear, specific questions. Upload comprehensive evidence. Link key evidence to relevant questions. Include framework documents for evaluation context. Provide direction when generating analysis. Check the evidence readiness assessment before proceeding.
Q: What if I have very little evidence? The AI can still generate analysis, but quality metrics will indicate low confidence. The value is identifying what gaps exist and what evidence you need.
Q: Can I generate analysis before uploading all my evidence? Yes, and it's often useful to do so early — the gap analysis will show you what's missing. Regenerate after adding more evidence for a more complete analysis.
Q: How do I handle contradictory evidence? Nquiry identifies contradictions in its analysis, but resolving them is your professional judgment. Provide direction that tells the AI how you want contradictions handled: "Evaluate the credibility of each account based on corroborating evidence and note which account is better supported."
About Reports
Q: Can I edit the AI-generated report sections? Yes. Report sections are drafts that you review and edit. The final report should reflect your conclusions.
Q: How should I cite AI assistance in my reports? Follow your organization's policies. Common approaches include noting in the methodology section that AI-assisted analysis tools were used, while maintaining that findings were verified and approved by qualified professionals.
Q: Do reports include citations to specific evidence? Yes. Report sections include inline citations that reference specific evidence items. You can click any citation to view the source evidence.
15. Glossary
Analysis — AI-generated evaluation of evidence against a specific question, topic, or investigation, including conclusion, confidence level, evidence evaluation, and quality metrics.
Analysis Chaining — The process where higher-level analyses (topic, summary) build upon lower-level analyses (question) rather than re-analyzing raw evidence from scratch.
Chunk — A segment of evidence text used for semantic search and AI processing, typically a few paragraphs in length.
Confidence Level — A rating indicating how well evidence supports a conclusion: Established, Probable, Possible, or Insufficient.
Context Budget — The allocation of the AI's context window across evidence, background documents, framework documents, and prior analyses.
Context Coverage — The percentage of your evidence items that were included in the AI's context for a specific analysis.
Coverage Score — Quality metric measuring whether the analysis addresses all component elements of the question.
Direction — Your guidance to the AI about how to approach analysis for a specific question.
Evidence Evaluation Framework — The structured criteria and rules the AI uses to evaluate evidence quality, derived from CIGIE, GAO, IIA, and ACFE standards.
Evidence Readiness — An assessment of whether your evidence collection is adequate for analysis, evaluated before the AI runs.
Faithfulness Score — Quality metric measuring whether claims in the analysis are supported by the evidence that was in context.
Finding Status — Your determination of whether a finding is Substantiated, Not Substantiated, or Inconclusive.
Focus Statement — The overarching purpose or scope of an inquiry.
Framework Document — A policy, regulation, or standard uploaded to your inquiry that the AI uses as evaluation criteria.
Link Divergence — When evidence you manually linked to a question differs from what the AI's semantic search identified as most relevant.
Quality Confidence Badge — A summary indicator (Established / Probable / Possible / Insufficient) combining faithfulness, coverage, and validation results.
Retrieval Quality — A measure of how relevant the evidence found by semantic search was to the analysis question.
Semantic Search — Finding content by meaning rather than exact keyword match, using mathematical representations of text meaning.
Topic — A logical grouping of related questions within an inquiry.
Document Version: 3.0 Last Updated: February 2026 For questions about this document, contact support@nquiry.ai