§ 04 The Existing Ecosystem
What already exists — and how QBiruni uses, learns from, and surpasses it.
Several powerful tools attack the scientific-literature problem. None of them solve it for quantum hardware engineering specifically. QBiruni does not compete with them — it is built on top of them, using each as the best available component for its role, while providing what none of them can: a domain-specific ontology, quantitative extraction with provenance, and tight integration with a running simulator.
Understanding this ecosystem is not optional. It is the fastest path to a working v0 — and the clearest map of where QBiruni's real moat lies.
Build on top of
PaperQA2 — FutureHouse
Open-source · State-of-the-art scientific QA with citation grounding
PaperQA2 is the current best-in-class for scientific question answering with provenance. It retrieves relevant passages, keeps citations attached to every answer fragment, and scores well on academic benchmarks. It is open-source and actively maintained.
→ Study its citation-grounding architecture before writing any extraction code. The problem of "every sentence must cite its source" is solved here.
→ Fork the retrieval-augmented QA loop as QBiruni's synthesis layer baseline.
→ Its weakness: no schema, no structured extraction. It returns prose, not typed rows. This is exactly where QBiruni takes over.
Use as infrastructure
Semantic Scholar API
Free API · 200M+ papers · Semantic search, citations, abstracts
Semantic Scholar is the primary paper database QBiruni should ingest from. It provides structured metadata (authors, venues, citation counts, semantic embeddings) and semantic search over 200M papers. It is not a competitor — it is the data layer.
→ Use the S2 API as primary ingestion source alongside the arXiv API. The two together cover > 98% of the relevant corpus.
→ Use S2's citation graph to surface papers that cite a key reference — essential for tracking how experimental claims evolve over time.
→ Filter by venue: PRX Quantum, Nature Physics, APL, PRL, arXiv:quant-ph, cond-mat.supr-con.
Use as filtering layer
arXiv-sanity-lite
Open-source · Relevance filtering for arXiv feeds
arXiv-sanity-lite (Karpathy's tool) provides a simple but effective relevance-ranking mechanism for arXiv papers using TF-IDF and SVMs. It is not semantic search — but it is fast, interpretable, and requires no GPU.
→ Use it as a coarse pre-filter before expensive LLM extraction. Papers ranked below threshold get metadata-only storage, not full extraction.
→ Fine-tune the term weights on quantum-hardware vocabulary: TLS, transmon, Josephson junction, tantalum, niobium, T₁, T₂, quasiparticle.
Benchmark against
Elicit + Consensus
Commercial · General-purpose scientific literature search
Elicit and Consensus are the best generic scientific-literature AI tools available today. They search papers, summarize claims, and extract some structured data. Every quantum hardware team is already aware of them or using them.
→ Use them as a benchmark baseline. Before claiming QBiruni is better, test the TLS oxide question on both and document where they fail. Concretely: they return prose summaries, not typed measurements; they hallucinate numbers; they have no concept of "device regime."
→ Their failure modes are QBiruni's differentiation story, made concrete with examples.
Learn from failure
Galactica — Meta AI
Discontinued · Large language model trained on scientific text
Meta's Galactica (2022) was trained on 48M scientific papers, textbooks, and knowledge bases. It was pulled from public access after 3 days because it confidently hallucinated scientific claims — plausible-sounding but wrong numbers, fabricated citations, invented results.
→ The lesson is structural: an LLM trained on science still hallucinates when not grounded in retrieved source text. QBiruni must enforce retrieval-first — never generate a number without a source paragraph in context.
→ Every extraction must store the verbatim source paragraph. This is not optional. It is the architectural lesson Galactica teaches.
Consider as extraction backbone
Llemma — EleutherAI
Open-source · Math and science LLM based on Code Llama
Llemma is a family of open-source models (7B and 34B) specifically trained on mathematical and scientific text. It outperforms general models on scientific reasoning tasks and runs locally — no API cost, no data leaving your environment.
→ Evaluate Llemma-34B as an extraction backbone for the structured ontology extraction step — potentially faster and more private than Claude API for high-volume extraction.
→ Trade-off: lower ceiling than Claude Sonnet on complex reasoning, but may be sufficient for schema-constrained extraction where the ontology does most of the work.
Study the interface pattern
SciSpace (Typeset)
Commercial · PDF reading and explanation for scientists
SciSpace allows researchers to upload PDFs and ask questions. It is widely used and has good UI/UX for the researcher workflow. It explains figures, tables, and equations in context.
→ Study its UI for the "explain this figure" and "what does this table mean" interactions — these are the interaction patterns QBiruni's researcher-facing interface should support in v1.
→ Its critical gap: it does not extract structured data, does not compare across papers, and has no simulation integration. QBiruni's value is exactly what SciSpace does not do.
Where QBiruni's moat actually lives
None of the above tools can do what QBiruni does: extract a T₁ value from a paper as a typed row — {device: tantalum transmon, substrate: sapphire, oxide_thickness: 3nm, T₁: 312µs ± 25, temp: 15mK, source: §3.2 para 4} — and automatically compare it against a QJosephson simulation output in the same schema. That tight ontological coupling between simulation and literature is the real moat. PaperQA2 does not know what a Josephson junction is. Semantic Scholar does not understand fabrication regimes. Elicit does not extract measurement uncertainty. QBiruni does all three — because it was built for this specific domain, with a schema designed jointly with QJosephson, by people who understand the physics.
The secondary moat is accumulation. Every paper that passes through QBiruni, every extraction that is audited and verified, every verdict that is checked against reality — this makes the database richer and the extractor more accurate. After 500 papers, QBiruni knows things about the quantum hardware measurement landscape that no human team has ever systematically compiled. That knowledge does not expire and is not easily replicated.
