How Wiki Generation Uses LLM Summaries for Community Pages in Code Review Graph

The wiki generator in code_review_graph does not call an LLM directly; instead, it embeds pre-computed LLM summaries stored in the description field of each community object.

The code-review-graph repository separates concerns between community analysis and documentation generation. The wiki system consumes LLM-produced descriptions that were generated earlier in the pipeline, making the documentation workflow efficient and reproducible.

How Community Descriptions Flow Into Wiki Pages

Step 1: LLM Summaries Are Produced During Community Discovery

The get_communities(store) function in code_review_graph/communities.py returns community objects as dictionaries. When communities are created, an external tool (such as a community-description generator) can invoke an LLM to analyze the code patterns within that community and store the resulting narrative in the "description" key.

This design allows flexibility—you can use any LLM provider or prompt engineering approach without modifying the wiki generator itself.

Step 2: The Wiki Generator Renders Pre-Computed Descriptions

In code_review_graph/wiki.py, the _generate_community_page(store, community) function handles page construction:

# From code_review_graph/wiki.py (lines 58-60)

description = community.get("description", "")
if description:
    md.write(f"## Overview\n\n{description}\n\n")

The generator performs these actions:

• Retrieves the description via community.get("description", "")
• If present, inserts it into the Overview section of the Markdown page
• Continues with additional sections: dependency tables, control flow diagrams, and cross-references

Step 3: The Final Markdown Page Contains the LLM Narrative

The complete workflow is: LLM → community description → wiki generator → Markdown page.

Complete Usage Example

Generate a wiki from an existing graph store:

from code_review_graph.wiki import generate_wiki

# store is already populated with communities and LLM descriptions

wiki_stats = generate_wiki(store, "/tmp/my_wiki")

print(wiki_stats)

# Output: {'pages_generated': 12, 'pages_updated': 3, 'pages_unchanged': 0}

Inspect a generated page to verify the LLM summary appears:

from pathlib import Path

page = Path("/tmp/my_wiki/data-processing.md").read_text()
print(page.splitlines()[:15])

# ['# Data Processing',

#  '',

#  '## Overview',

#  '',

#  'This module provides utilities for handling CSV and JSON parsing...',

#  ...]

Manually inject a description for testing or customization:

from code_review_graph.communities import get_communities

communities = get_communities(store)
communities[0]["description"] = (
    "Handles all I/O operations with retry logic and async support."
)

# Subsequent wiki generation will embed this text

Key Files in the Wiki Generation Pipeline

File	Responsibility
code_review_graph/wiki.py	Core wiki generation; embeds community["description"] into the Overview section of each page
code_review_graph/communities.py	Constructs community dictionaries; description field is populated by external tools
code_review_graph/tools/docs.py (example)	Illustrative location where an LLM tool might generate descriptions before wiki creation

Why This Architecture Matters

Decoupled design allows you to:

• Swap LLM providers without touching wiki code
• Cache expensive LLM calls (descriptions are stored, not regenerated)
• Generate wikis offline or in air-gapped environments
• Version-control community descriptions independently of the generator

Summary

• The wiki generator reads LLM summaries from community["description"] rather than calling an LLM
• Descriptions are inserted into the Overview section of each community page in code_review_graph/wiki.py
• Community objects originate in code_review_graph/communities.py with pre-populated metadata
• This pipeline pattern keeps documentation generation fast, deterministic, and independent of LLM availability

Frequently Asked Questions

Where does the LLM actually get called in the codebase?

The repository does not mandate a specific LLM call location. Typically, a separate tool in code_review_graph/tools/ (such as docs.py) or an external pipeline invokes the LLM to produce community descriptions before generate_wiki() runs. The wiki system itself is LLM-agnostic.

Can I generate wiki pages without any LLM summaries?

Yes. The community.get("description", "") call returns an empty string when no description exists, and the Overview section is omitted. The page still contains dependency tables, flow diagrams, and structural information derived from static code analysis.

How do I customize which LLM generates the descriptions?

Implement a custom tool that calls your preferred LLM and attaches the result to each community dictionary before passing it to generate_wiki(). The wiki generator requires only that the "description" key exists; it does not inspect how the value was produced.

Is the description regenerated every time I run generate_wiki()?

No. Descriptions are persisted in the community objects within the graph store. They are written once during community discovery and read repeatedly during wiki generation. To refresh descriptions, you must re-run the community discovery pipeline with your LLM tool.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}