A certified, traceable company brain for your organization — one your AI agents and your team can query.
Every fact scored 0–1 for certainty, traced to its source, and honest about what it doesn't know yet.
Try building that in a weekend.
Not a folder of files. Not a vector index. A living knowledge layer that turns raw documents into scored, sourced facts — organized into domains, kept correct over time, and served to whoever asks: a human, or one of your agents.
↩ And it learns. When the brain can't answer confidently, it doesn't bluff — it records the gap, an expert fills it, and the original question gets re-answered.
Take the interactive tour →Most AI tools bluff when they're unsure. SciPot refuses to. When the brain can't answer with confidence, it records a Knowledge Gap — a tracked work item that turns every “I don't know” into the brain getting smarter exactly where it was weak.
An assistant that admits when it doesn't know, then learns — instead of confidently making something up. Trust you can build decisions on.
A deterministic is_knowledge_gap signal on every answer, plus a gap lifecycle you drive via the API. No more silent failures to debug.
Your RAG retrieves. SciPot knows. You built retrieval — but retrieval is not knowledge.
A POT is not a vector store. It's a cognitive environment where raw documents transform into validated, interconnected knowledge with explicit certainty scores.
Every fact gets a POT Score. Every relationship is typed and weighted. Contradictions are detected and surfaced. Human curators govern the constitution while automated agents handle extraction at scale.
Upload a document. SciPot handles the rest.
Five jobs a real brain has to do — from the atom of knowledge to the loop that keeps it learning.
LLM-powered extraction turns documents into atomic, scored knowledge units with full provenance and quality audit.
Every fact carries an explicit certainty score from 0.0 to 1.0. Filter queries by confidence. No more guesswork.
Every fact is auditable: who curated it, who resolved conflicts, and the exact document and fragment it came from.
Typed, weighted links: supports, contradicts, derives_from, refines, is_part_of — a graph, not a pile.
One domain = one brain-region. Each POT keeps its mission, scope, and facts coherent instead of dissolving into one pile.
A small set of foundational axioms locked at POT Score 1.0 — the ground truth every new fact is checked against.
Agents extract and score at scale. Human experts govern the constitution and resolve conflicts — with an audit trail.
When two facts disagree, the brain notices and flags it for a human to resolve — instead of silently averaging them.
Supporting evidence boosts confidence, contradictions lower it, and scores decay over time as knowledge ages.
Each question is classified and routed across cache, vector, graph, and document strategies — fast where it can be, deep where it must be.
Answers whose wording tracks certainty, with source attribution and explicit warnings about low-confidence or conflicting facts.
Tracks what it can't answer, routes it to an expert, then re-answers automatically. The brain gets smarter exactly where it was weak.
The POT Score makes knowledge quality visible and actionable. Query for facts you can stake decisions on, or explore hypotheses with lower-confidence knowledge.
Scores propagate through the knowledge graph: supporting evidence boosts confidence, derivations inherit from premises with decay, and contradictions are flagged for human review.
| 1.0 | Constitution | Foundational axioms, set by human curators only |
| 0.85 | Verified | Multiple sources or explicit human validation |
| 0.50 | Extracted | Extracted with reasonable confidence from sources |
| 0.30 | Inferred | System-inferred from combining other facts |
| 0.00 | Pending | Awaiting verification or insufficient evidence |
The same certified knowledge, served two ways. Ask in plain language and get a finished answer, or pull structured facts your agent can reason over — both with full provenance and certainty.
A natural-language answer whose wording tracks the evidence — “It is established that…” for verified facts, “Preliminary analysis suggests…” when it's inferred. Confident only when it should be.
Ranked facts, a relationship subgraph, and an epistemic briefing that tells your agent how to read the certainty landscape — so it keeps full reasoning control instead of trusting a black box.
import httpx
response = httpx.post(
"https://api.scipot.ai/pots/{pot_id}/synthesize",
headers={"Authorization": "Bearer ik_live_..."},
json={
"query": "What are the key risk factors?",
"max_facts": 10
}
)
result = response.json()
print(result["answer"]) # Synthesized response
print(result["facts_used"]) # Facts with pot_score & provenance
print(result["contradictions_found"]) # Detected conflicts
print(result["epistemic_briefing"]) # Certainty guidance
print(result["strategy_used"]) # "cag", "rag", etc.We spent two centuries dividing labor. The harder problem now is dividing judgment — and you can't lay it along an assembly line. So you give it a place: bounded spaces where what's known is checked, certified with an honest confidence label, and reused a thousand times instead of reinvented. That's a Knowledge Pot.
SciPot turns any domain into a knowledge-rich environment. Here are some of the highest-value applications developers are building today.
Ingest financial statements, contracts, and regulatory filings into a POT. Detect contradictions across documents and synthesize risk assessments with calibrated certainty scores.
Finance & M&AMaintain a living POT of regulations—FDA, GDPR, SOX—that detects when new rules conflict with existing practices and synthesizes compliance impact for specific products.
Compliance & LegalOne POT per project: architecture decisions, postmortems, runbooks. Certainty scores distinguish firm decisions from casual suggestions. Knowledge compounds with every interaction.
Developer ToolsIngest papers, clinical trials, and datasets. Score propagation through the knowledge graph reflects how real scientific evidence works—not all findings carry equal weight.
Pharma & BiotechUse a POT as the operational brain for agent swarms. Agents log decisions and outcomes; scores distinguish validated patterns from hypotheses. Dual curation lets humans correct what agents learn.
Agent InfrastructureUse a POT as canonical ground truth to evaluate agent outputs in production. Detect when an agent contradicts validated facts, overclaims confidence, or introduces unverified knowledge—turning contradiction detection into a runtime guardrail.
AI ObservabilityUsage-based plans for teams shipping on our cloud — or a self-hosted license when the data has to stay in yours.
Tokens cover all operations: ingestion, fact extraction, edge detection, queries, and synthesis.
Service never cuts off abruptly — soft warnings and gradual throttling on overage.
SciPot deployed inside your own AWS VPC — your documents and knowledge never leave your account. Billed as a flat license, not per token, so cost is predictable no matter how hard your agents work. Includes SSO, audit logging, dedicated support, and an SLA.
KB2B is a full-stack knowledge management application built entirely on SciPot's API. Persistent conversations, validated knowledge, domain-specific expertise — a working proof of what you can build.
Visit kb2b.app →Choose your entry point.
Questions? hello@scipot.ai