Paper Ingest Normalizer#

Convert raw literature inputs into standardized records safe for project memory, paper databases, and downstream synthesis pipelines.

Input#

One of the following is required:

• pdf_path — local path to PDF file
• url — link to paper/article
• raw_text — extracted or pasted text
• metadata_blob — existing metadata dict

Plus:

• project_id — required for any writeback
• source_type — one of: pdf, doi, url, text, metadata
• optional tags — list of strings for categorization

Output Schema#

Return a structured object:

text

title: string
authors: string[] | null
year: number | null
source: string          # journal, conference, preprint, etc.
doi_or_url: string | null
project_id: string
paper_type: string      # experimental, theoretical, review, etc.
material_system: string | null   # e.g. "钙钛矿太阳能电池", " graphene FET"
device_type: string | null       # e.g. "FTO/glass", "flexible substrate"
key_variables: string[] | null   # independent variables studied
key_metrics: string[] | null     # measured outcomes (PCE, mobility, etc.)
core_findings: string            # 2-3 sentence neutral summary
claimed_mechanism: string | null
limitations: string | null
normalized_summary: string       # 1-2 paragraph structured summary
uncertain_fields: string[] | null  # fields that could not be verified
writeback_ready: boolean        # true only if key identity fields present
writeback_payload: object        # the record to write into project memory

Rules#

1. Never write into project memory without project_id. Ask if not provided.
2. Separate direct observations from claimed interpretations. Mark inference vs. direct extraction.
3. Preserve uncertainty. Use null for missing fields; list in uncertain_fields.
4. Do not invent missing bibliographic fields. Don't hallucinate authors, year, etc.
5. Do not over-claim. Keep core_findings and normalized_summary grounded in what the text actually says.
6. Never conflate abstract with findings. The abstract states intentions; findings are what the data supports.
7. If writeback_ready = false, list explicitly which fields are missing and why.

PDF Extraction#

For PDFs, use the summarize skill or pdfplumber/ PyMuPDF to extract text before processing.

Workflow#

1. Identify source type — determine which input field is populated
2. Extract raw content — PDF text, URL content, or use provided raw text
3. Parse bibliographic fields — title, authors, year, source, DOI
4. Identify research content — material system, device type, variables, metrics
5. Distill findings — separate what was measured from what was claimed
6. Assemble writeback_payload — structured record matching the schema above
7. Assess completeness — set writeback_ready based on presence of key identity fields

Failure Handling#

If parsing is incomplete:

• Return partial structured output with all successfully extracted fields
• Populate uncertain_fields with the list of fields that could not be determined
• Set writeback_ready = false when title, authors, or year are missing

Cross-Reference#

For synthesis after normalization, see the research skill for paper synthesis workflows.