Skip to content

AI-Assisted Features Overview

Askalot integrates AI assistance throughout the survey research workflow, helping researchers design, execute, and analyze surveys more efficiently while maintaining formal mathematical rigor.

The AI-Assisted Workflow

AI capabilities are embedded at each stage of the survey lifecycle:

flowchart LR
    subgraph design["1. Design"]
        d1[Research Brief] --> d2[Research Assistant]
        d2 --> d3[Structured Brief]
        d3 --> d4[Designer]
        d4 --> d5[SMT Validation]
    end

    subgraph campaign["2. Campaign"]
        c1[Wizard Guidance] --> c2[Strategy Design]
        c2 --> c3[Pool Generation]
        c3 --> c4[Launch]
    end

    subgraph execute["3. Execute"]
        e1[Real Respondents] --> e2[Survey Completion]
        e3[Persona Simulation] --> e2
    end

    subgraph analyze["4. Analyze"]
        a1[Bronze Dataset] --> a2[Quality Metrics]
        a2 --> a3[AI Analyst]
        a3 --> a4[Structured Report]
    end

    design --> campaign
    campaign --> execute
    execute --> analyze

Feature Summary

Feature Stage Purpose Key Capabilities
Questionnaire Generation Design Transform research briefs into validated QML Designer agent with research/planning/writing/review sub-agents, semantic indexing, SMT validation
Campaign Management Campaign Guide campaign setup and execution Wizard flow, chat interface, sampling strategy design
Response Generation Execute Simulate campaigns with synthetic data Persona profiles, realistic responses, pipeline testing
Result Analysis Analyze Evaluate data quality and generate reports Sample representativeness, response quality metrics, AI analyst reports
AI Functions & Model Options All Reference: which model runs each AI function Embedding, layout detection, OCR (planned), extraction, re-ranking, cluster labeling, reasoning — what each requires and what you can choose

The AI Agents

Askalot's AI assistance is delivered by four top-level agents, each owning one stage of the research lifecycle. Three of them orchestrate specialist sub-agents — focused helpers they delegate to for a single, well-defined job — while the Respondent works on its own. Every agent runs on Claude and draws on Askalot's distilled survey-methodology skills, reaching for the peer-reviewed methodology library when a question needs citable evidence.

Agent Stage What it does Specialist sub-agents
Designer Design Turns a research brief and uploaded documents into a formally verified QML questionnaire — researching the source material, planning the structure, writing each section, and reviewing the finished design. Research Assistant — analyzes reference documents and drafts the structured Research Brief
QML Planner — decomposes the brief into an ordered chapter plan
QML Writer — writes the formally verified QML one chapter at a time
Design Reviewer — audits the saved, error-free questionnaire's design quality and returns a scorecard
Manager Campaign Runs the campaign lifecycle: sampling-strategy design, respondent-pool generation, campaign creation, and fielding oversight. Field Supervisor — monitors live data collection (response rates, segment coverage, quality flags) and recommends quota, targeting, or timeline adjustments
Analyst Analyze Evaluates collected results and synthesizes a unified report covering both statistical data quality and whether the data answers the original research goals. Quality Analyst — statistical data-quality assessment (representativeness, weighting effectiveness, response quality)
Research Evaluator — checks results against the Research Brief's research questions, success criteria, and requirements
Respondent Execute Completes surveys as a simulated respondent, staying consistent with a demographic persona and the study's interviewer brief — used for testing, validation, and demos. (none — works independently)

Each top-level agent decides when to hand off: the Designer delegates research, planning, writing, and design review to its four specialists; the Manager passes live-fielding questions to the Field Supervisor; the Analyst splits evaluation into a data-quality pass and a research-goal pass, then combines the two. Delegating to focused sub-agents keeps each agent's reasoning sharp and lets the specialists carry deep, task-specific skills.

AI-Assisted Questionnaire Generation

Stage: Design | Tool: Armiger

Transform research documents and briefs into formally verified QML questionnaires with the Designer agent and its specialist sub-agents.

Working in one continuous conversation, the Designer delegates to a Research Assistant that analyzes uploaded documents (PDFs, specs, regulations) and drafts a structured Research Brief, a QML Planner that turns the brief into an ordered chapter plan, a QML Writer that writes formally verified QML one chapter at a time — validating each against the Z3 SMT solver — and a Design Reviewer that audits the finished questionnaire's design quality.

Learn more

AI-Assisted Campaign Management

Stage: Campaign | Tool: Targetor

Two complementary interfaces for campaign setup and management.

Campaign Wizard

Guided multi-step flow for beginners:

  1. Campaign Design - Name, questionnaire selection, mode
  2. Target Audience - Demographics, sample size
  3. Sampling Strategy - Stratification factors, distributions
  4. Respondent Pool - Generate and preview pool quality
  5. Review & Launch - Summary, validation, deployment

Chat Interface

On-demand conversational interface for power users:

  • "Merge these two pools to get more respondents"
  • "Show me which interviewers have the lowest completion rates"
  • "Create a new pool with only respondents aged 25-34"
  • "Compare response rates between campaign versions"

Learn more

Agentic Response Generation

Stage: Execute | Tool: Portor MCP (mass_fill_surveys + get_task_status)

Generate synthetic survey data for testing, validation, and demo purposes. mass_fill_surveys defers a background task executed by Portor's own worker — the tool returns a task_id immediately and callers poll get_task_status until terminal. The tool offers four distribution strategies to balance speed vs. response quality:

Distribution Strategies

Strategy Engine Speed Quality Best For
realistic Weighted random Instant Medium Default — demographically influenced responses
random Uniform random Instant Low Stress testing, edge case discovery
stratified Strata-matched Instant Medium Quota-balanced samples
llm Claude Haiku ~1s/question High Realistic synthetic data, demos, presentations

llm mode produces the highest quality synthetic data: each question is answered by Claude Haiku with the full persona profile, respondent demographics, and accumulated Q&A history for intra-survey consistency. A respondent who answers "unemployed" won't describe a workplace later.

Rule-based modes (realistic, random, stratified) use weighted random selection based on persona traits — faster for high-volume pipeline testing where response quality matters less.

Learn more

AI-Assisted Result Analysis

Stage: Analyze | Tool: Balansor

Evaluate survey data quality across two dimensions and generate methodology-grounded reports.

Two Quality Dimensions

Dimension Metrics Question Answered
Sample Representativeness RMSE, MAE, Chi-Square, Max Deviation, Quality Score Does the sample match the target population?
Response Quality Normalized Entropy, Straightlining, Cronbach's Alpha, Acquiescence Bias Are responses reliable and informative?

AI Analyst Agent

The analyst agent runs quality assessment tools via MCP, gathers campaign context, and produces a structured report:

  1. Executive Summary — overall fitness for purpose
  2. Sample Representativeness — per-factor breakdown with specific numbers
  3. Weighting Assessment — Bronze vs Silver improvement analysis
  4. Key Findings — data-driven observations
  5. Recommendations — prioritized, actionable next steps

The agent is grounded in established survey methodology (AAPOR, Kish, Groves, Krosnick, ESOMAR) and interprets statistical metrics in practical research context.

Learn more

Grounding in Peer-Reviewed Methodology

Askalot's Designer, Manager, and Analyst agents consult a shared library of 24 peer-reviewed survey-methodology books and papers (Dillman, Krosnick, Groves, Tourangeau, Bethlehem, Heeringa, Schouten, Fowler, and others) covering the full research process — design, sampling, fielding, analysis, weighting — when a question goes beyond their built-in skills or when a recommendation needs citable evidence.

The library is a graph-aware knowledge base built from Docling-parsed markdown of each paper, stored in the LightRAG vector + graph backend, and retrievable in five modes (keyword, entity-centric, concept-centric, hybrid, and mixed). Agents cite the paper, year, and section when they rely on it, so researchers can verify the claim directly.

Agent Topics it reaches for the library
Designer Validity and reliability (Taherdoost, Aithal), questionnaire logic and skip patterns (Fagan & Greenberg, Elliott, Feeney & Feeney, Schiopu-Kratina, Manski & Molinari), question wording and response scales (Krosnick, Bradburn, Fowler)
Manager Sampling design, stratification, cluster sampling, design effects and weighting fundamentals (Heeringa, Bethlehem, Schouten), adaptive survey design
Analyst Total Survey Error, nonresponse bias, raking and post-stratification, response-quality frameworks

The library is consulted on demand — agents default to their distilled skills for routine questions to keep latency low, and only reach for the full corpus when the customer asks for a citation or when the question sits outside the skill's scope.

Integration Across Stages

The AI features work together seamlessly:

Document → Data Flow

Research Documents
    ↓ [AI Questionnaire Generation]
QML Questionnaire
    ↓ [AI Campaign Management]
Campaign with Sampling Strategy
    ↓ [Agentic Response Generation - optional]
Survey Responses (real or simulated)
    ↓ [AI Result Analysis]
Insights and Reports

Shared Context

AI assistants share context across the workflow:

  • The Designer's document analysis informs its QML generation
  • Questionnaire structure informs sampling strategy recommendations
  • Campaign demographics guide persona selection for simulation
  • Survey responses feed directly into analysis tools
  • Insights can trigger questionnaire refinements (closing the loop)

Project-Scoped Isolation

All AI-indexed documents are scoped by the active project, ensuring that research materials stay within team boundaries. In non-private organizations, each user's AI sessions are automatically scoped to their default project. This isolation carries through the entire workflow — from document indexing to the Designer agent's MCP tool access when it writes and validates QML.

AI Provider

All AI agent tasks in Askalot run on Anthropic's Claude models, which the platform reaches through either of two gateways:

Gateway Models Best For
Anthropic API Claude Opus, Sonnet, Haiku Pay-as-you-go key or personal Claude subscription
AWS Bedrock Claude Opus, Sonnet, Haiku Enterprise deployments that need AWS IAM, VPC, and data-residency controls

Customers bring their own API credentials — no vendor lock-in on the account level, and Bedrock lets enterprises keep all traffic inside their own AWS account. You choose your provider explicitly under Profile Settings → AI Models.

"AI" here is not one model doing everything: embedding, document layout analysis and passage re-ranking run on Askalot's own hardware, while the generative work runs on Claude through your chosen provider. Which of those you can configure, and what happens when a credential is missing, is mapped function by function in AI Functions & Model Options.

Explicit choice, no silent fallback

Each AI capability uses the provider you select — there is no hidden automatic switching, and no platform-owned key behind the scenes. If you choose a provider but its credential is missing or expired, the feature stops and asks you to fix it in Profile Settings (or re-authenticate, for an expired AWS session) rather than quietly running on a different provider, on someone else's account, or returning a degraded result that looks real. This keeps you in control of where your data and spend go.

Local Inference

The platform runs its own inference service for document embedding (the BGE-M3 model that powers semantic indexing in Questionnaire Generation), document layout analysis, and optional cross-encoder re-ranking of retrieved passages — so that content is not sent to a model provider at all. Embedding is fixed to the local model: indexing and search must use the same model, so it is not a configurable choice. Everything generative — questionnaire design, campaign management, response simulation, result analysis, knowledge-graph extraction, query keyword extraction, and optional cluster labeling — runs on Claude via the Anthropic API or AWS Bedrock.

Local models sleep when idle and wake on first use

To stay quiet and power-efficient when no one is indexing or searching, the local inference service powers down between requests. The first request after an idle gap wakes it — your action still completes, but it may take a few extra seconds. Askalot makes this visible rather than leaving you guessing: a "Local models" indicator in the top navigation bar (and in the code editor's status bar) shows whether the local models are ready, waking, or asleep, and document indexing and Designer chat show a brief "waking local models" note while they warm up. AI agents calling the platform's tools get the same signal — a tool result notes when its call included a cold-start wait. Nothing fails and nothing asks you to retry; the wait is simply surfaced.

Getting Started