Incorrect Citation Format

The Problem

AI provides sources in inconsistent formats, with wrong URLs, missing page numbers, or citations that don't match retrieved content.

Symptoms

• ❌ Citations point to wrong documents

• ❌ "Source: Unknown" despite using context

• ❌ Inconsistent citation style

• ❌ Made-up source references

• ❌ Cannot verify claims

Real-World Example

AI response: "The rate limit is 1000/hour [Source: API Documentation, page 15]"

Problems:
→ Document has no page numbers (web page)
→ Claim is from chunk_id: api_v2_limits_003
→ Citation format doesn't match actual source structure

User cannot verify the claim

---

Deep Technical Analysis

Citation Generation Methods

Inline Metadata:

Pass source info in context:
"[doc_id:api_guide_v2, title:'API Limits', url:'/docs/api#limits']
The rate limit is 1000 requests per hour."

LLM must extract and format:
→ Often loses metadata
→ Or formats inconsistently

Post-Processing:

Extract citations after generation:
1. Identify factual claims in response
2. Match to source chunks
3. Append citations

More reliable but complex

---

How to Solve

Implement structured citation format in system prompt + pass chunk IDs inline with content + use post-processing to verify citations + provide citation template ("According to [doc]...") + include direct URLs in metadata. See Citation Management.