Citation Attribution Assembly

Background

RAG outputs cite their sources inline (e.g. ... [1] ... [3]). The UI parses these markers out of the answer text to render footnotes / inline source links. Markers are 1-indexed in the answer, 0-indexed in the source-id list, and should be deduplicated in first-seen order.

Problem statement

Implement extract_citations(answer). Walk answer; for every [<int>] marker (whitespace tolerated inside the brackets), subtract 1 to convert to 0-indexed. Track a seen set; emit each citation only the first time it appears. Return a list[int] in first-seen order.

Input

• answer — str (or None/empty).

Output

Returns list[int] — the 0-indexed citation IDs in first-seen order.

Examples

Example 1 — simple list

Input:  "This is [1] and also [3]"
Output: [0, 2]

Example 2 — dedup preserves order

Input:  "[2] then [1] then [2] again"
Output: [1, 0]

Example 3 — whitespace inside brackets tolerated

Input:  "see [ 5 ]"
Output: [4]

Example 4 — no citations

Input:  "plain text"
Output: []

Constraints

• Citations are integers wrapped in [...] (regex \[\s*(\d+)\s*\]).
• Non-integer brackets ([ref]) are skipped.
• 1-indexed in the answer text; 0-indexed in the output.

Notes

• Verification. Calling code should check that each emitted index is within the source list (0 <= idx < len(sources)); a model can hallucinate [99].
• Numbered vs anchor. Some RAG stacks use anchor IDs ([doc_42]) instead of integers. Same kernel, different regex.