Hybrid 4-layer answering engine

Stop searching your literature.
Ask it questions.

H4Graph turns thousands of papers into a knowledge graph your team can interrogate — and answers with citations, including the connections you didn’t know were there.

Every answer cited Your corpus, never trained on API-first
h4 ask — “which methods outperform transformers on long documents?”
semanticmatchPaper A2023Method XoutperformsPaper Bcites Acontradicts2025 · 3 hops
[vector] 12 entry passages by meaning · 0.3s
[graph] CITES → EVALUATED_ON → CONTRADICTS · 3 hops
[relational] 9 claims pinned to source text + DOI
[agent] contradiction found (2025) — answer cited ✓
The H4 engine

Hybrid 4-layer answering

Vector-only RAG answers what sounds similar. Real research questions need structure, ground truth, and judgment. Each layer covers the others' blind spot.

LAYER 01

Vector

Qdrant · embeddings

Finds passages by meaning, not keywords. Your question matches evidence phrased in entirely different vocabulary.

Blind spot it covers: exact-word search misses paraphrased science.
LAYER 02

Graph

Neo4j · knowledge graph

Traverses citations, methods, datasets, and concepts — multi-hop. Finds the contradiction published two years later that shares no words with your question.

Blind spot it covers: similarity search dies one hop deep.
LAYER 03

Relational

PostgreSQL · system of record

Grounds every claim: exact source text, DOIs, dates, and licence rights. Turns candidate evidence into evidence you may legally show — and precisely cite.

Blind spot it covers: LLMs invent citations. This layer can't.
LAYER 04

Agent

Claude · goal-driven planning

Plans which registered traversals to run, notices gaps, re-queries, and synthesises a structured answer — every statement wired to its source.

Blind spot it covers: pipelines can't adapt. An agent replans.

Any single layer fails a real research question. H4Graph is the claim that four together don’t — compressed into seven characters.

Why not just RAG?

The question every “chat with your PDFs” tool gets wrong

“Which methods outperform transformers on long documents — and has anyone contradicted those results?

Vector-only RAG1 HOP

  • Finds chunks that sound like your question — and stops there
  • The contradicting 2025 paper shares no vocabulary → invisible
  • Answers the first half, silently misses the second
  • Citations generated, not guaranteed

H4GraphMULTI-HOP

  • Vector finds the entry points by meaning
  • Graph follows CITES → EVALUATED_ON → CONTRADICTS, three hops out
  • Relational layer pins every claim to source text, DOI and date
  • Agent notices the gap, re-queries, answers both halves — cited
  • Ambiguous entity? It clarifies rather than guesses
Governance

“Cited” isn’t a prompt. It’s the architecture.

Most tools ask their model to cite sources and hope it complies. H4Graph's answering layer is built so it cannot emit an untraceable value.

MECHANISM 01

Registered traversals only

The agent never writes queries. It plans over a catalog of read-only, depth-limited query templates — each one profiled, index-backed, and tested like code. No injection surface, no runaway traversals.

Entity doesn't resolve? It asks — clarifies rather than guesses.
MECHANISM 02

Per-field lineage

Every value in an answer records the template, bound parameters, graph nodes, and source passage that produced it — written to an append-only store the application cannot rewrite.

Insert-only by database privilege, not by policy.
MECHANISM 03

Reproducible on demand

Any answer replays from its trace: template + parameters + graph version. An auditor walks from a sentence in an answer to the exact source passage and the exact query that retrieved it.

Built for reviews that ask “show me.”
From PDFs to answers

Four steps. Zero reading lists.

01 · DISCOVER

Bring the literature in

Search OpenAlex, Semantic Scholar, arXiv, CrossRef, PubMed — or bring your own licences: Scopus, Embase, IEEE, Springer, Wiley. Internal reports come straight from Snowflake or Databricks.

02 · INGEST

Agents build the graph

Papers are chunked, embedded, and mined for entities and relationships — authors, methods, datasets, citations — into one knowledge graph.

03 · ASK

Interrogate the corpus

Natural-language questions, answered across all four layers. Multi-hop, cross-paper, quota-metered.

04 · TRUST

Every claim, cited

Answers link back to the exact source passage. Publishable provenance, not plausible prose.

The corpus, visible

Browse your literature as a living graph

Every paper, method, concept, and author your corpus contains — connected, searchable, and inspectable. Click any node to read exactly where it came from and which query retrieved it.

EVALUATED_ONCONTRADICTSLongformer (2020)BigBird (2020)S4 (2021)Hyena (2023)Mamba (2023)Zhao et al. (2025)Sparse attentionState-space modelsLong-document QATransformersA. GuT. Dao
Z
PAPER
Zhao et al. (2025)
doi10.1234/zhao.2025
year2025
statusCONTRADICTS
source neo4j
node_id p:1088
template contradiction_scan@2
graph_ver 2026-07-01T09:00Z
12 nodes · 16 edges · auto layout · 2d

live demo · search a method, drag a node, click for provenance — the amber edge is the contradiction

API-first

Wire it into your pipelines

A versioned public API, generated SDKs, and signed webhooks. Your LIMS, internal tools, and data pipelines get the same cited answers your team does.

  • /api/v1 — stable by contract. OpenAPI-pinned, additive-only changes, Pact-verified on every merge.
  • Webhooks, not polling. paper.ingested, answer.completed — HMAC-signed, retried.
  • M2M auth that fits. API keys for simple integrations, OAuth2 client-credentials with scopes for enterprise.
  • Idempotent by design. Safe retries on every mutating call.
quickstart.pypip install h4graph
from h4graph import H4Graph

h4 = H4Graph(api_key="h4_live_…")

# ingest — batched, webhook on completion
h4.papers.submit(doi="10.1234/example.2025")

# ask — hybrid 4-layer answer, cited
answer = h4.questions.ask(
    "Which methods outperform transformers
     on long documents? Any contradictions?"
)

print(answer.text)
for c in answer.citations:
    print(c.doi, c.passage, c.hops)  # provenance, always
Pricing

Start free. Scale with your corpus.

Free

Trial the graph
$0
  • 10 papers
  • 25 questions / month
  • Open-source connectors
  • Cited answers
Start free

Researcher

Individual academics
$29/mo
  • 200 papers
  • 500 questions / month
  • OpenAlex · Semantic Scholar · arXiv · CrossRef · PubMed
  • Priority email support
Choose Researcher
MOST POPULAR

Lab

Research groups
$99/mo
  • 2,000 papers · 5 seats
  • 3,000 questions / month
  • Shared team graph
  • BYOL connectors (Scopus, IEEE…)
  • Prepaid credit packs
Choose Lab

Enterprise

Pharma & R&D orgs
$499/mo+
  • Unlimited papers · custom quotas
  • Full BYOL catalog + internal docs
  • Snowflake & Databricks doc connectors
  • Public API · SSO · scoped keys
  • Regional data residency
  • Annual commit discounts
Talk to us

Heavy month? Prepaid credit packs from $25 — volume bonuses up to +25%, no surprise bills, ever.

Built to pass compliance review

Never trained on

Your corpus stays yours. Nothing you ingest is used to train models — architecturally, not just contractually.

Auditable provenance

Every answer traces to source text, DOI, date, and licence rights — evidence your compliance team can replay.

Residency on your terms

Pin storage and inference to the region you choose. Your data never leaves the geography your review requires.

License-aware by design

Ingestion stores only what each source's licence permits — metadata, abstract, or full text, enforced per connector. Never relicensed.

The connections are already in your literature.
Start finding them.

Ten papers, cited answers, five minutes to your first “I didn’t know that.”

Start free