Source Attribution

Every source in an answer carries a percentage. This page explains what that number measures, how Miso calculates it, and what it does not mean.

Together they answer one question: of all the material behind this answer, how much came from this source?

For the rest of the answer payload, see Answer Response Fields.

Every entry in sources carries two extra fields:

Together they answer one question: of all the material behind this answer, how much came from this source?

%%{init: {'theme':'base','themeVariables':{
'fontFamily':'IBM Plex Sans, ui-sans-serif, system-ui','fontSize':'14px',
'lineColor':'#9aa3c0','clusterBkg':'transparent','clusterBorder':'#d3d8e8',
'titleColor':'#8b93ad'
},'flowchart':{'curve':'basis','nodeSpacing':20,'rankSpacing':60,'padding':12}}}%%
flowchart LR
  Q["`**The question**`"]
  S["`Search your catalog`"]
  C10["`10 candidates
  sent to the model`"]
  C3["`**3 cited**
  they become sources`"]
  C7["`7 not cited
  ignored`"]
  PCT["`Percentages
  over the cited 3 only`"]

  Q --> S --> C10
  C10 --> C3 --> PCT
  C10 --> C7

  classDef q fill:#f6f7fa,stroke:#c9cfe0,color:#454c69
  classDef work fill:#e8ecfb,stroke:#b9c2ea,color:#22307a
  classDef cited fill:#f4effc,stroke:#cdbcf0,color:#40277a
  classDef drop fill:#f6f7fa,stroke:#d8d8e2,color:#8b93ad
  class Q q
  class S,C10 work
  class C3,PCT cited
  class C7 drop

How an answer is built

Miso does not answer from memory. It reads your content first, in three steps:

1. Search. Miso finds the passages that relate to the question.
2. Send. Miso sends the best passages to the model. Each one is a candidate.
3. Cite. The model writes the answer, and cites the candidates it used.

A candidate the model cites becomes a source. The others are dropped.

Sent is not the same as cited

This distinction controls the whole calculation.

Send 10 candidates, and the model cites 3. Those 3 become sources, their percentages add up to 100, and the other 7 add nothing.

The calculation

1. Take the cited sources. Ignore every other candidate.
2. For each one, measure the passage Miso sent to the model.
3. Add the lengths together.
4. Divide each length by the total, then multiply by 100.

See the worked example below for a real one.

A worked example

The example below uses the Wikinews example catalogue. The question:

What did Wikinews report about the Linux operating system?

Miso answered from three articles:

Wikinews reported extensively on the major shift in the CentOS ecosystem following Red Hat's December 2020 announcement to move focus away from CentOS Linux in favor of CentOS Stream…

The sources array beside it, trimmed to the fields that matter here:

[
  {
    "product_id": "wikinews-2911293",
    "title": "Gregory Kurtzer discusses plans for Rocky Linux with Wikinews",
    "date": "2020-12-18T00:00:00+00:00",
    "_attribution_length": 9588,
    "_attribution_length_percentage": 59.28
  },
  {
    "product_id": "wikinews-2911067",
    "title": "Red Hat to move focus away from CentOS in favour of Stream",
    "date": "2020-12-14T00:00:00+00:00",
    "_attribution_length": 4845,
    "_attribution_length_percentage": 29.96
  },
  {
    "product_id": "wikinews-2911401",
    "title": "Wikinews 2020: An 'Original reporting' year in review",
    "date": "2021-01-01T00:00:00+00:00",
    "_attribution_length": 1741,
    "_attribution_length_percentage": 10.76
  }
]

Check it yourself. The three passages add up to 9588 + 4845 + 1741 = 16,174 characters, and each percentage is that source's share:

Rounding to two decimal places means the total can read 100.01 or 99.99 on another answer. Do not test for exactly 100.

Note what the numbers do not say. The interview scored 59.28% because it is a long interview, not because it settled the question. The year-in-review article scored 10.76% because Miso sent a short passage from it. The percentage follows the length of the passage, not how much of it the model used.

Which text is measured

Three lengths exist for one source. Miso measures the middle one:

Miso removes the layout markup before it measures.

How to read it

The percentage measures supplied material, not importance. The model can read a long passage and take one fact from it. Three consequences follow:

• Length shapes the number. A longer passage scores higher than a short one.
• A short source can carry the key fact. Its percentage stays small, because the percentage follows the length of the passage.
• The number of citations does not change it. One citation or five, the passage length is the same.

Take a reader who asks for the date of a court decision:

Both figures are correct. They describe volume. The feature supplied most of the material, and the notice supplied the decisive fact. Read the percentage with the answer text and the citation markers to see how each source was used.

In short