Skip to content

AI Functions & Model Options

"AI" in Askalot is not one model doing everything. It is a set of distinct functions, each chosen for a specific job, each with its own requirement and its own degree of choice. Some are platform-fixed because a per-user choice would break correctness. Others are yours to select, because they spend your budget and touch your data.

This page is the complete map: what each function does, what it needs to work, what you can change, and what happens when something is missing.

At a glance

Function What it does Runs on Your choice?
Embedding Turns text into vectors for semantic search and clustering Askalot's own inference server (BGE-M3) No — platform-fixed
Layout detection Finds headings, paragraphs, tables and reading order in a PDF Askalot's own GPU (Docling) No — infrastructure
OCR Reads text out of scanned pages and images Not enabled yet Planned
Knowledge-graph extraction Pulls entities and relations out of your documents Claude (your Direct LLM slot) Yes
Re-ranking Re-scores retrieved passages for relevance Askalot's own inference server (BGE reranker) On/off per query
Query keyword extraction Works out what to look for in the knowledge graph Claude (your Direct LLM slot) Yes
Cluster labeling Names the themes found in open-ended answers Keyword statistics, or Claude (your Direct LLM slot) Yes — method and model
Reasoning (the agents) Designer, Manager, Analyst, Respondent Claude (your Reasoning slot) Yes
Document summaries The index of what your uploaded documents cover Claude (your Reasoning slot, cheap tier) Follows Reasoning
Respondent simulation Answers surveys as a persona (llm mode) Claude Haiku, via your Reasoning provider Provider only

Everything marked "Claude" runs through your own Anthropic API key, Claude subscription, or AWS Bedrock account — see Choosing your models.

Retrieval functions

These four turn an uploaded document into something an agent can search. They run in sequence: parse the page, split it into chunks, vectorize the chunks, build a knowledge graph over them.

Embedding — platform-fixed

What it does. Converts every document chunk and every search query into a 1024-dimensional vector, so "how do I calculate a design effect" can find a passage that never uses those words.
Model BGE-M3 (multilingual, 8K context)
Where it runs Askalot's own GPU inference server — your text is not sent to a third party for this step
Requirement The inference server must be reachable. It sleeps when idle and wakes on first use (a few extra seconds).
Options None. The slot exists in Profile Settings but is fixed and read-only.

Why you cannot change the embedding model

Search only works if your documents were indexed with the same model the query uses. Two different models produce vectors that are not comparable — switching would not degrade search, it would silently return nothing useful from everything indexed before the switch. So the choice is made once, at the platform level, and kept.

This also keeps your documents' vectors on Askalot's own hardware rather than on a third-party embedding API.

Layout detection — platform infrastructure

What it does. Before text can be chunked it has to be read correctly: headings identified, columns un-interleaved, tables recovered as tables rather than as scrambled runs of numbers.
Model Docling layout analysis + TableFormer (table structure recovery)
Where it runs Askalot's own GPU
Requirement Applies to PDFs. Plain text, Markdown and CSV skip this step entirely.
Options None — it is infrastructure, not a model preference

Why this is not a setting

CPU and GPU produce byte-identical output here — the GPU is roughly four times faster, and that is the only difference. Parsed documents are also cached and shared, so a document one person uploads is not re-parsed for the next person. A per-user preference would be quietly ignored whenever someone else's parse populated the cache first, which is worse than having no setting at all.

OCR — planned

What it does (when it lands). Reads text out of scanned pages, photographed documents, and images embedded in a PDF.
Status Not enabled. OCR is switched off in the current document pipeline.
What this means today A PDF with a real text layer indexes normally. A scanned PDF — pages that are images of text — yields little or no text, so the document indexes as effectively empty rather than failing with an error.
How to work around it Run the scan through OCR yourself before uploading, or upload a text-layer version.
Planned options Whether OCR runs at all, and the language hint, are the two knobs under consideration. No dates yet.

Check scanned uploads

If you upload a scan and the Designer cannot find anything in it, this is almost certainly why — not a retrieval failure. Open the document's indexing status: a scanned page set produces a very small chunk count.

Re-ranking — on/off per query

What it does. After the graph and vector search return candidate passages, a cross-encoder re-reads each candidate against your actual question and re-scores it. It is more accurate than vector similarity because it looks at the query and the passage together rather than comparing two independent summaries of them.
Model BGE-reranker-v2-m3 (cross-encoder)
Where it runs Askalot's own GPU inference server
Default Off
Options On or off, per query

Why it defaults to off

Quality is not free here. Scoring 40 full-length candidates costs about 8 seconds per query — real quality, real latency. For most retrieval the graph + vector ranking is already good enough, so the extra wait is not worth paying on every question. Turn it on when precision matters more than speed.

Generative functions

These call a Claude model. They spend your credits and send the relevant text to whichever provider you selected, so each one is tied to a slot you control.

Knowledge-graph extraction — your Direct LLM slot

What it does. Reads each chunk of your document and pulls out entities and the relations between them, building the graph that makes retrieval conceptual rather than merely keyword-based. This is the most token-hungry AI step in the platform — a long PDF is many chunks, and each one is a model call.
Runs on Claude, via your Direct LLM slot (Anthropic API or AWS Bedrock)
Requirement A configured Direct LLM slot with a working credential
Options Provider and model
If unconfigured Indexing fails with a clear error telling you to finish setup

Why extraction is billed to you

Extraction is the single largest AI cost in document indexing. Attributing it to the person who uploaded the document is what makes that cost visible and attributable, instead of pooling it invisibly into platform overhead. The same principle runs through every generative function on this page: the account you chose is the account that pays, and you can always see which one that is.

Query keyword extraction — your Direct LLM slot

What it does. When you ask a question, a fast, cheap model call works out which entities and concepts to look for in the knowledge graph before retrieval runs.
Runs on Claude, via your Direct LLM slot
Requirement Same slot as extraction above
Options Provider and model — a small, fast model is the right choice here

Cluster labeling — method, then model

What it does. Open-ended text answers are clustered into themes (see Semantic Clustering). Each cluster then needs a name — and you choose how that name is produced.

Method Runs on Speed Result
Keyword statistics (default) Askalot, no network call Instant Top distinguishing keywords, e.g. "delivery, late, refund"
LLM Claude, via your Direct LLM slot ~1 s per cluster A natural-language theme name, e.g. "Delayed deliveries and refund requests"
Default Keyword statistics — it always works, needs no credential, and costs nothing
Requirement for LLM mode A configured Direct LLM slot with a working credential
If the credential is rejected The coding job fails and says so

Why a rejected credential fails the job

The failure mode this prevents is the quiet one. If labeling fell back to generic placeholders — Theme 0, Theme 1, Theme 2 — the dataset would still be marked ready, the columns would still be there, and nothing would look wrong until someone tried to read the analysis. An expired AWS session in the middle of a run is exactly the case that would trigger it. So the job stops and tells you, and you re-run it after fixing the credential.

Reasoning — the four agents

What it does. Everything the Designer, Manager, Analyst and Respondent do: reading your brief, planning a questionnaire, writing and validating QML, running a campaign, interpreting quality metrics.
Runs on Claude, via your Reasoning slot
Options Anthropic API key, personal Claude subscription, or AWS Bedrock
Tiers High / medium / low — each stage of work uses the tier that suits it, all on the provider you chose
If unconfigured The agent stops and asks you to finish setup — it never silently switches provider

Local models are deliberately excluded here: the agent framework requires Claude-compatible models. That is a capability constraint, not a judgement about local model quality.

Document summaries — follows Reasoning

What it does. Produces the short index of what each uploaded document covers, so the Designer knows what material exists and what to go looking for. One cheap, single-turn call per document plus one to combine them.
Runs on Your Reasoning slot, cheapest tier
Options Follows your Reasoning choice — nothing separate to configure

Respondent simulation — provider only

What it does. In llm distribution mode, answers each survey question as a demographic persona, carrying the conversation history forward so a respondent who said "unemployed" does not describe a workplace three questions later.
Runs on Claude Haiku, through your Reasoning slot's provider
Options Provider only — the model is fixed to the fast, cheap tier because the job is high-volume and short-form
Alternative The three rule-based modes (realistic, random, stratified) need no AI at all

Choosing your models

Three settings, under Profile Settings → AI Models. They are independent — each covers a different way of invoking a model, and choosing one does not constrain the others.

Slot Covers You can choose
Embedding Semantic indexing and search Fixed to the local model — shown but not editable
Direct LLM Knowledge-graph extraction, query keyword extraction, cluster labeling Anthropic API, or AWS Bedrock
Reasoning The four agents, document summaries, respondent simulation Anthropic API, Claude subscription, or AWS Bedrock

Credentials are entered separately under Profile Settings → AI Providers. An organization administrator can set organization-wide credentials; your personal credential takes precedence when both exist.

No shared key, no silent fallback

Askalot never runs your AI work on a platform-owned credential. A model slot resolves your personal credential first, then your organization's — and stops there. If neither is set, the feature tells you to configure it rather than quietly running on someone else's account.

The same rule covers expiry. If your AWS session has lapsed, you get a re-authenticate prompt, not a silent switch to a different provider or a degraded result that looks like a real one.

Where your data goes

Embedding, layout detection and re-ranking run entirely on Askalot's own hardware — that content is not sent to a model provider. The generative functions send the relevant text to the provider you selected: Anthropic directly, or AWS Bedrock inside your own AWS account if you need IAM, VPC or data-residency control.

When something is missing

Askalot fails loudly and specifically, by design. The table below is the map from symptom to cause.

Symptom Cause Fix
Indexing fails with a credential error Direct LLM slot has no working credential Add a key under Profile Settings → AI Providers, or re-authenticate AWS
Coding job fails on LLM labeling Direct LLM credential missing or rejected Same as above, or switch that recipe to keyword labeling
Agent asks you to finish setup Reasoning slot has no working credential Choose a provider and add its credential
Indexing succeeds but the document seems empty Scanned PDF with no text layer — OCR is not enabled yet Upload a text-layer version, or OCR it first
First action after a quiet period is slow Local inference models were asleep and are waking Nothing — the action completes; the navigation bar's Local models indicator shows the state
Retrieval feels imprecise Re-ranking is off by default Turn it on for that query, accepting the added latency