> ## Documentation Index
> Fetch the complete documentation index at: https://wuweism.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Analyzing research papers with PDF upload

> How to attach PDFs to Causal Chat queries and use PDF Synthesis to extract structured evidence from multiple papers at once.

Wu-Weism can analyze research papers directly. You can attach a PDF to any Causal Chat message — the extracted evidence then grounds the response in the paper's actual data — or use the dedicated **PDF Synthesis** tool to cross-analyze up to six papers simultaneously.

Both surfaces extract structured numeric evidence, generate governed claims, and record everything in the Claim Ledger.

## Two ways to use PDFs

<Tabs>
  <Tab title="Causal Chat attachment">
    Attach a single PDF to a Causal Chat message. The PDF's content is extracted and used to inform the causal response to your question. Best for:

    * Grounding a specific causal question in a paper's reported data
    * Checking whether a paper's findings support or contradict a claim
    * Extracting key statistics while also running causal analysis
  </Tab>

  <Tab title="PDF Synthesis (/pdf-synthesis)">
    Upload up to six PDFs at once and optionally specify a research focus. Wu-Weism reconciles evidence across all documents, flags contradictions, and synthesizes a unified causal picture. Best for:

    * Systematic literature review
    * Identifying where papers agree and disagree on a causal mechanism
    * Generating a set of governed claims from a body of evidence
  </Tab>
</Tabs>

## Analyzing a single paper in Causal Chat

<Steps>
  <Step title="Open a Causal Chat conversation">
    Navigate to **/chat** and start a new conversation or continue an existing one.
  </Step>

  <Step title="Click the attachment button">
    Click the **paperclip icon** (attachment button) in the message input bar. A file picker will open. Select your PDF file.

    The PDF name will appear as an attachment chip next to the message input. You can attach one PDF per message.
  </Step>

  <Step title="Type your question">
    Write your causal question in the message input as you normally would. The question will be analyzed in the context of the attached paper.

    ```text Example question with PDF attached theme={null}
    What causal claims about IL-6 and fibrosis progression does this paper support?
    ```

    ```text Another example theme={null}
    Does the methodology in this paper meet the identifiability requirements 
    for a Rung 2 intervention claim?
    ```
  </Step>

  <Step title="Read the three-section output">
    When a PDF is attached, the response follows a structured **three-section output contract**:

    ***

    **Section 1: All explicit numbers with context**

    Every numeric value extracted from the paper, grouped by category:

    | Category          | Examples                                                  |
    | ----------------- | --------------------------------------------------------- |
    | Potential metrics | Effect sizes, p-values, confidence intervals, odds ratios |
    | Structural        | Sample sizes, group counts, time points, dosages          |
    | Bibliographic     | Publication year, journal impact factor if present        |
    | Citation years    | Years of cited works                                      |
    | Reference indices | Reference numbers linked to cited claims                  |

    ***

    **Section 2: Claim-eligible numerics**

    A filtered subset of Section 1 containing only the values that are precise enough and contextually clear enough to anchor a governed claim. Values with ambiguous units, missing context, or contradictory definitions elsewhere in the paper are excluded from this section.

    ***

    **Section 3: Three claims with uncertainty labels**

    Three governed claims derived from the paper's evidence. Each claim carries an **evidence class** label:

    | Evidence class                  | Meaning                                                                                                    |
    | ------------------------------- | ---------------------------------------------------------------------------------------------------------- |
    | `bibliographic/structural only` | The claim is supported only by citation references or study design information — no direct numeric metrics |
    | `mixed`                         | The claim combines structural evidence with some numeric data, but the metrics are indirect or aggregated  |
    | `metric-bearing`                | The claim is directly supported by explicit numeric measurements from the paper                            |

    ```text Example Section 3 output theme={null}
    [Claim 1 — metric-bearing]
    Anti-inflammatory treatment reduced IL-6 levels by 38% (95% CI: 24–51%) 
    in the intervention arm vs. control at 12 weeks.
    Uncertainty: metric-bearing | Confidence: 0.79

    [Claim 2 — mixed]
    Fibrosis progression was attenuated in the treatment group, consistent 
    with reduced IL-6 signaling, though direct fibrosis measurements were 
    not stratified by IL-6 quartile.
    Uncertainty: mixed | Confidence: 0.52

    [Claim 3 — bibliographic/structural only]
    The study design is consistent with prior RCT methodology for cytokine 
    intervention trials (citations: [12], [14], [17]).
    Uncertainty: bibliographic/structural only | Confidence: 0.41
    ```

    All three claims are automatically recorded in the Claim Ledger.
  </Step>
</Steps>

<Warning>
  Wu-Weism extracts text from PDFs using the document's text layer. **Scanned PDFs** — images of printed pages with no embedded text — will have significantly lower extraction quality. Key statistics may be missed, misread, or absent from Sections 1 and 2. If you are working with scanned documents, use OCR software to create a searchable PDF before uploading. Claims derived from scanned PDFs will be labeled with lower confidence scores.
</Warning>

## Using PDF Synthesis for multi-paper analysis

PDF Synthesis is a dedicated research surface at **/pdf-synthesis** designed for analyzing multiple papers at once.

<Steps>
  <Step title="Navigate to /pdf-synthesis">
    Open **PDF Synthesis** from the workbench navigation.
  </Step>

  <Step title="Upload your PDFs">
    Drag and drop up to **six PDF files** into the upload area, or click **Select files** to use the file picker. Each file appears in the upload queue with its name and size.
  </Step>

  <Step title="Optionally specify a research focus">
    In the **Research focus** field, enter a brief description of what you are investigating. This helps Wu-Weism weight evidence and frame claims around your specific question rather than the papers' full scope.

    ```text Example research focus theme={null}
    Effect of anti-inflammatory interventions on IL-6 in patients with 
    chronic kidney disease
    ```

    <Tip>
      Providing a research focus consistently improves synthesis quality. Without it, Wu-Weism treats each paper equally and generates claims that reflect the papers' own framings, which may not align with your specific research question. Even a one-sentence focus statement makes a meaningful difference.
    </Tip>
  </Step>

  <Step title="Review synthesized output">
    The synthesis produces:

    * **Numeric evidence summary**: Aggregated metrics across all papers, including table counts, trusted table counts (tables where headers and data are reliably parsed), and total data point counts per paper.
    * **Cross-paper claim comparison**: Where multiple papers address the same causal relationship, Wu-Weism surfaces agreement and contradiction.
    * **Governed claims set**: Claims derived from the full corpus, each attributed to the source paper(s) and labeled with an evidence class.
    * **Reconciliation notes**: Where papers conflict on effect direction or magnitude, the synthesis explicitly flags the disagreement rather than averaging it away.
  </Step>
</Steps>

## What gets extracted

When Wu-Weism processes a PDF — whether in Causal Chat or PDF Synthesis — it extracts:

| Extracted field        | Description                                                          |
| ---------------------- | -------------------------------------------------------------------- |
| Numeric evidence       | All numeric values with surrounding context                          |
| Table counts           | Number of tables identified in the document                          |
| Trusted table counts   | Tables where structure is reliably parsed (headers and rows matched) |
| Data point counts      | Total discrete numeric data points extracted                         |
| Claims                 | Candidate claims derived from evidence                               |
| Bibliographic metadata | Where present: authors, year, journal                                |

## Claims and the Claim Ledger

Every claim produced from a PDF analysis — whether in Causal Chat or PDF Synthesis — is automatically recorded in the [Claim Ledger](/governance/claims) with:

* Source paper name
* Section of the paper the claim derives from
* Evidence class label
* Confidence score
* Session and timestamp

You can review, annotate, challenge, and export these claims from **/claims** at any time.

## Limitations

* **Text layer required**: PDFs must have an embedded text layer. Scanned-only documents have significantly degraded extraction.
* **Complex table layouts**: Multi-level headers, merged cells, and sideways tables may not parse correctly into the trusted table count.
* **Supplementary materials**: If supplementary data is in a separate file, upload it as a separate document in PDF Synthesis or attach it in a second Causal Chat message.
* **Non-English papers**: Extraction and claim generation quality is highest for English-language documents.

## Next steps

* Combine PDF evidence with intervention questions: [Running interventions](/guides/interventions)
* Understand how claims are governed and audited: [Claim Ledger](/concepts/claim-ledger)
* Run a full multi-paper synthesis with the Hybrid Synthesis tool: [Hybrid Synthesis](/workbench/hybrid-synthesis)