Footnotes and References Lost

The Problem

Footnotes, endnotes, and citations are separated from their reference markers during chunking, losing critical context and academic/legal citations.

Symptoms

• ❌ Reference markers ([1], *, †) appear without footnotes

• ❌ Footnotes separated from main text

• ❌ "See note 5" but note 5 not in chunk

• ❌ Academic citations incomplete

• ❌ Legal references missing source

Real-World Example

Original document:

The API rate limit¹ is 1000 requests per hour for free users².

────────────────────────
¹ Rate limits reset at midnight UTC
² Enterprise plans have higher limits

Chunk boundary falls between text and footnotes ↓

Chunk 1:
"The API rate limit¹ is 1000 requests per hour for free users²."

Chunk 2:
"¹ Rate limits reset at midnight UTC
² Enterprise plans have higher limits"

User sees Chunk 1: "What does ¹ mean?"
AI cannot resolve reference (footnote in different chunk)

---

Deep Technical Analysis

Footnote Types and Formats

Different notation systems:

Numbering Systems:

Numeric: [1], [2], [3] or ¹, ², ³
Alphabetic: [a], [b], [c]
Symbolic: *, †, ‡, §, ||, ¶
Roman: [i], [ii], [iii]

Placement Variations:

Bottom of page (traditional):
Main text with marker¹
────────────────────
¹ Footnote content

End of section:
Main text with marker¹
[Section continues...]
Notes:
¹ Footnote content

End of document (endnotes):
Main text with marker¹
[Many pages later...]
Endnotes:
¹ Footnote content

Detection Challenges:

Superscript numbers:
→ Could be footnote: "value¹"
→ Or exponent: "x²"
→ Or chemical: "CO₂"

Context needed to distinguish

Square brackets:
→ Could be footnote: "research [1] shows..."
→ Or citation: "see [Smith, 2020]"
→ Or array notation: "array[1]"
→ Or just brackets: "[optional] parameter"

Marker-to-Note Matching

Associating references with definitions:

Matching Algorithm:

1. Scan document for markers: ¹, ², ³
2. Scan document for footnote definitions:
   → Look for "¹ " at start of line
   → Or "──" separator followed by notes
3. Match markers to definitions by number/symbol
4. Store associations

Challenges:
→ Multiple numbering systems in same doc
→ Nested footnotes (footnotes referencing footnotes)
→ Reused numbers across chapters
→ Non-contiguous numbering (1, 2, 5, 7 - some missing)

The Multiple Reference Problem:

Main text:
"This concept¹ was explored by several researchers¹."

Same footnote referenced twice
→ Footnote 1 appears once at bottom
→ Two markers in text
→ Must link both to same footnote

Cross-Chapter Footnotes:

Chapter 1 footnotes: 1-15
Chapter 2 footnotes: 1-12 (numbering restarts!)

Marker "2" in Chapter 2 ≠ Marker "2" in Chapter 1
→ Must track context (chapter/section)
→ Avoid mixing footnotes across chapters

Academic Citations

Scholarly documents use formal citations:

Citation Formats:

IEEE: [1], [2], [3] (numeric)
APA: (Smith, 2020) (author-year)
MLA: (Smith 24) (author-page)
Chicago: Superscript¹ (note-based)
Harvard: (Smith 2020, p.15) (author-year-page)

Inline vs Bibliography:

Inline citation:
"Previous research [1] demonstrated..."

Bibliography (separate section):
[1] Smith, J. (2020). Title of Paper. Journal Name, 15(3), 123-145.

Chunking challenge:
→ Citation [1] in main text
→ Full reference in bibliography (different location)
→ May be in different chunk entirely

LLM sees [1] but can't resolve to full citation

Citation Clustering:

Multiple citations together:
"This is well documented [1,2,3,5-8,12]."

Represents:
→ 9 separate citations (1,2,3,5,6,7,8,12)
→ Must expand ranges (5-8)
→ Link all to bibliography

If chunk contains this line:
→ Need ALL 9 bibliography entries
→ But they may span multiple pages
→ Impractical to include all

Legal Citations

Legal documents have specific citation requirements:

Legal Citation Format:

Case law: Smith v. Jones, 123 F.3d 456 (9th Cir. 2020)
Statute: 42 U.S.C. § 1983
Regulation: 17 C.F.R. § 240.10b-5

Components:
→ Case name
→ Reporter volume & page
→ Court
→ Year

The String Citation:

Legal writing uses "string citations":
"This principle is established. See Smith v. Jones, 123 F.3d 456, 460 (9th Cir. 2020); Doe v. Roe, 789 F.2d 123, 125 (2d Cir. 2019); Johnson v. Williams, 456 F.Supp. 789 (S.D.N.Y. 2018)."

Single sentence with 3 citations
→ Must keep together
→ Splitting loses context
→ "See Smith v. Jones" alone is incomplete

Abbreviated Citations:

First reference (full):
"Smith v. Jones, 123 F.3d 456 (9th Cir. 2020)"

Later references (short):
"Smith, 123 F.3d at 460"
or just: "Id. at 461" (same case as previous)

"Id." depends on previous citation
→ Must track citation history
→ If chunk starts mid-document: "Id." unresolvable

Footnote Content Length

Footnotes vary from brief to extensive:

Short Footnotes:

Main text: "The API¹ supports JSON."
Footnote: ¹ Application Programming Interface

Brief definition: 5-10 words
→ Easy to include with main text

Long Footnotes:

Main text: "The algorithm¹ is efficient."
Footnote: ¹ The algorithm is based on dynamic programming principles first introduced by Bellman (1957) and later refined by Dijkstra (1959). Modern implementations typically use a priority queue for efficiency. For a comprehensive treatment of the theoretical foundations, see Cormen et al. (2009), Chapter 24. Note that the worst-case time complexity is O(n²) for dense graphs but can be reduced to O(n log n) with appropriate data structures. [200 words...]

Footnote longer than main text!
→ Including footnote inflates chunk size
→ Excluding footnote loses critical detail

The Inclusion Decision:

Options:
1. Always include footnotes in chunk
   → Pros: Complete context
   → Cons: Very large chunks, repetition

2. Never include footnotes
   → Pros: Smaller chunks
   → Cons: Incomplete information

3. Include short footnotes (<50 words)
   → Pros: Balance
   → Cons: Arbitrary threshold, inconsistent

4. Include footnotes as separate chunks with back-references
   → Pros: Modular
   → Cons: Complex linking required

Inline Notes vs Margin Notes

Different annotation styles:

Inline Parenthetical:

"The result (as shown in Figure 3) demonstrates..."

Not a footnote, but similar
→ Interruptive aside
→ Could be moved to footnote
→ But author chose inline

Should chunk preserve parenthetical positioning?
Or normalize: "The result demonstrates... See Figure 3."

Margin Notes:

Document layout:
┌────────────────┬──────────┐
│ Main text here │ Note: Im-│
│ continues with │ portant  │
│ more content.  │ detail!  │
└────────────────┴──────────┘

Margin note parallel to main text
→ Not clearly "after" any specific paragraph
→ Extraction order ambiguous
→ Associate with nearby text? Which paragraph?

Reference Loops and Nested Notes

Complex referencing structures:

Footnote Referencing Footnote:

Main text: "The concept¹ is fundamental."
Footnote 1: "See also related work²"
Footnote 2: "Smith (2020) provides comprehensive review."

Multi-level reference chain
→ Resolving ¹ requires ²
→ Deep linking required

Circular References:

Footnote A: "See Footnote B for details"
Footnote B: "As mentioned in Footnote A..."

Circular dependency
→ Cannot resolve independently
→ Must include both together

Embedding and Retrieval Impact

Footnotes affect semantic search:

Footnote Content in Embeddings:

Query: "API rate limit reset time"

Option 1: Embed main text only
"The API rate limit is 1000 requests per hour."
→ Doesn't match query (no "reset" mentioned)

Option 2: Embed main text + footnotes
"The API rate limit is 1000 requests per hour. Rate limits reset at midnight UTC."
→ Matches query! Footnote has answer

Conclusion: Must include footnotes for complete semantic coverage

Citation Noise:

Academic paper full of citations:
"This approach [1,2,3] outperforms baseline [4,5] with significant improvements [6,7,8,9]."

Embedding includes: [1,2,3,4,5,6,7,8,9]
→ Citation numbers are noise
→ No semantic value
→ Dilute actual content signal

Should strip [1-9] before embedding?
→ But then lose traceability
→ Cannot cite sources

---

How to Solve

Detect footnote markers (superscripts, brackets) + match to footnote definitions at page/section end + inline short footnotes (<50 words) directly + link long footnotes as metadata + resolve "Id." and abbreviated citations + strip citation brackets from embeddings but store separately. See Footnote Handling.