Citation Index — Stable URL-to-Number Map

Borrowed from STORM's

StormArticle.url_to_unified_index

+ deepagents

research_agent/prompts.py

citation consolidation. Keeps citations stable across report sections and subagent outputs.

State

citation_index:
  entries:
    # URL (canonicalized) → entry
    "https://arxiv.org/abs/2401.12345":
      number: 1
      title: "First paper title"
      first_seen_section: "introduction"
      used_in_sections: ["introduction", "background", "method"]
      source_channel: "search-arxiv"
    "https://github.com/foo/bar/issues/42":
      number: 2
      title: "Issue: memory leak"
      first_seen_section: "implementation-notes"
      used_in_sections: ["implementation-notes"]
      source_channel: "search-github-issues"
  next_number: 3

URL Canonicalization Rules

Before indexing, normalize the URL:

1. Strip tracking params:

   utm_*

   ,

   gclid

   ,

   fbclid

   ,

   ref=

   ,

   source=

   .
2. Strip fragments (

   #section-1

   ) unless the URL is an anchor-addressed resource (e.g. docs page).
3. Lowercase host but keep path case.
4. Collapse multiple slashes (

   //

   →

   /

   ) except for the scheme.
5. Platform-specific:
   • YouTube:

     watch?v=ID

     form is canonical; strip

     list=

     playlist params.
   • arXiv: strip version suffix for de-dup (

     v2

     →

     ""

     ); keep title for reference.
   • GitHub issues: canonicalize

     githubissues.com/X/Y/Z

     to

     github.com/X/Y/issues/Z

     .

Different URL canonicalizations → different citation numbers. A URL that points to the same resource but failed canonicalization is a bug; log it and merge.

Write Path

When a subagent / section outputs evidence:

for ev in evidence:
    url = canonicalize(ev.url)
    if url not in index.entries:
        index.entries[url] = Entry(
            number=index.next_number,
            title=ev.title,
            first_seen_section=current_section,
            used_in_sections=[current_section],
            source_channel=ev.source_channel,
        )
        index.next_number += 1
    else:
        if current_section not in index.entries[url].used_in_sections:
            index.entries[url].used_in_sections.append(current_section)

Read Path

When rendering a section body, replace inline citation markers with the assigned number:

• Input from runtime AI:

  ... as shown in [ref: arxiv.org/abs/2401.12345] ...

• After indexing:

  ... as shown in [1] ...

• References section at end:

  [1] Title. URL.

Merge Rule (Subagent Consolidation)

When merging N subagent outputs:

1. Collect all evidence URLs across all subagents.
2. Canonicalize and dedupe.
3. Assign numbers in stable order (first-seen wins; use subagent order as tie-break).
4. Rewrite each subagent's inline

   [ref: ...]

   tags to the final numbers.

Failure Modes

• Malformed URL in evidence — log and skip; don't crash the whole index.
• Duplicate title with different URLs (e.g. "paper.pdf" published on two mirrors) — keep both; this is user judgment, not ours to merge without data.

When to Use

• Final report synthesis (always).
• Multi-subagent merge (always).
• Single-section simple answer — skip, overkill.

Cost

Cheap — mostly bookkeeping. No LLM calls unless resolving ambiguous cases (e.g. "are these two URLs the same resource?"). Runtime AI typically never needs to consult an LLM for citation indexing.

MCP Tool Usage

Full citation workflow using MCP tools:

# Create an index for this research session
idx = citation_create()
index_id = idx["index_id"]

# Add URLs as you collect evidence (idempotent — same URL always gets same number)
num1 = citation_add(index_id=index_id, url="https://arxiv.org/abs/2501.12345",
                    title="RAG Survey 2026", source="arxiv")["citation_number"]
num2 = citation_add(index_id=index_id, url="https://github.com/user/repo",
                    title="rag-toolkit", source="github")["citation_number"]

# Merge citations from a parallel delegate_subtask result
# (if you ran delegate_subtask with a separate citation_create per subtask)
citation_merge(target_id=index_id, source_id=other_index_id)

# Export as Markdown reference list
refs = citation_export(index_id=index_id)["markdown"]
# refs = "[1] RAG Survey 2026 — arxiv (https://arxiv.org/abs/2501.12345)\n[2] rag-toolkit ..."

Use

[1]

,

[2]

inline citations in your report body, then append

refs

at the end.

Interactions

• Fed by → all channel skills +

  delegate-subtask

  output.
• Feeds →

  synthesize-knowledge

  (which produces the

  [1] [2]

  report).
• Feeds →

  evaluate-delivery

  (which can spot "[5] is referenced but not in the index" bugs).

Quality Bar

• Evidence items have non-empty title and url.
• No crash on empty or malformed API response.
• Source channel field matches the channel name.