Built in collaboration with the UGA Department of Physics & Astronomy
Raman and infrared spectroscopy papers report peak-assignment data — the frequency and mode-assignment tables spectroscopists rely on — with no standardized reporting format. Different papers use different notation conventions (Mulliken, Herzberg, Wilson), inconsistent table structures, and metadata buried in prose rather than clean fields. Manually transcribing decades of this literature into a usable, structurally consistent dataset is slow, error-prone, and doesn't scale to building a literature-wide spectroscopy database.
Any real solution needs to:
- Handle heterogeneous formatting across papers written decades apart, by different research groups, using different notation traditions
- Produce structurally consistent, normalized output rather than one-off, ad hoc transcriptions
- Keep a human in the loop for verification, rather than blindly trusting fully automated extraction
SpectraGuru is a multi-stage pipeline that takes a scientific PDF and turns it into a normalized, structured dataset. It reads the paper, figures out which tables actually contain peak-assignment data, pulls out each individual peak measurement, and enriches every row with standardized chemistry and vibrational-mode metadata — all validated against a rigorously designed data schema built in collaboration with the SpectraGuru research team. The result is a growing, literature-wide dataset that's consistent enough to analyze at scale, and detailed enough to eventually serve as training or benchmark data for downstream machine learning work (e.g., predicting molecular structure from a spectrum).
| Layer | Choice | Why |
|---|---|---|
| OCR / PDF ingestion | Datalab/Marker hosted OCR API | Offloads a genuinely hard problem — robust PDF table-structure extraction — to a specialized external service rather than reinventing PDF parsing |
| Extraction engine | Google Gemini (gemini-3-flash-preview) |
Chosen for strong structured-output support and multimodal PDF understanding |
| Output constraints | Pydantic v2 models (extra="forbid") passed as JSON Schema to the LLM's structured-output API |
Correct, idiomatic use of schema-constrained generation — the model can only emit outputs matching the defined schema, rather than relying on prompt instructions and free-text JSON scraping |
| Verification strategy | Two-pass self-consistency check between the content-validation stage and the table-isolation stage | A lightweight, prompt-level technique for reducing output variance — the model is asked to cross-validate its own answer against its own prior output rather than trusting a single pass |
| Persistence (current) | Checkpointed flat JSON files after every stage | Makes each stage idempotent and resumable — a failure late in a long document doesn't waste the cost and latency of earlier LLM calls |
| Data model | A 79-parameter normalization schema spanning provenance, instrument/acquisition parameters, analyte chemistry (SMILES, CAS numbers), and vibrational-mode taxonomy | Real domain modeling, developed and iterated with a research collaborator — this is the project's deepest and most differentiated asset |
Data flow: PDF → OCR'd markdown → LLM-validated candidate tables → cross-checked Raman-specific subset → extracted peak rows (strict schema validation) → enriched vibrational-mode metadata (+ parallel reference-list extraction) → checkpointed JSON.
A note on scope: this is a research-infrastructure tool for a single research group's literature-review workflow, not a consumer product — the architecture is intentionally linear and single-document at this stage, rather than over-built with infrastructure the current scale doesn't need.
| Component | Notes |
|---|---|
| PDF → Markdown OCR (Datalab/Marker) | Handles table-structure-preserving conversion |
| Content validation (candidate table detection) | Schema-constrained Gemini call |
| Table isolation + self-consistency cross-check | Two-pass verification between stages |
Row extraction (PeakRow schema) |
Strict Pydantic validation, extra="forbid" |
| Vibrational mode enrichment | 10+ normalized fields per row (notation, symmetry, SMILES, etc.) |
| Reference log extraction | Runs alongside row extraction |
| 79-parameter domain schema | Designed and iterated with a research collaborator |
| Hand-labeled gold-standard dataset | Built as ground truth for evaluation |
Note on metrics: the pipeline includes a hand-built gold-standard evaluation dataset, but it is not yet wired into an automated scoring script. Field-level extraction accuracy, per-stage latency, and token cost are being actively measured and will be added here once the evaluation harness is complete.