Agent Knowledge Packs¶

Turn any documentation into an AI agent skill.

Knowledge Packs are self-contained, domain-specific knowledge graphs that install as agent skills in Claude Code. Each pack bundles curated content from documentation into a local graph database with vector search — when the agent needs domain expertise, it retrieves grounded context from the pack instead of relying on training data.

# Build a pack from documentation URLs
echo "y" | uv run python scripts/build_go_pack.py

# Package and install as a skill
cd data/packs && tar -czf go-expert.tar.gz go-expert
wikigr pack install go-expert.tar.gz

# Now just ask Claude Code about Go — the skill activates automatically

Why Knowledge Packs?¶

LLMs have three limitations that packs solve:

Limitation	The Pack Fix
Training cutoff — can't know about new APIs and versions	Packs ingest current documentation
Depth gaps — knows basics but misses implementation details	Packs contain full docs with section-level granularity
Hallucination — plausible but wrong answers on niche topics	Pack answers trace back to specific source articles

Evaluation: 99% Accuracy¶

48 packs evaluated across 2,716 questions (judged by Claude Opus):

Metric	Training (Claude alone)	With Knowledge Pack
Accuracy	91.7%	99%
Pack wins	—	38 of 48 (79%)

See full results.

The Pack Lifecycle¶

1. Describe — Choose a domain and curate URLs¶

Create data/packs/my-pack/urls.txt with 50-80 documentation URLs.

2. Build — Fetch, extract, embed, store¶

echo "y" | uv run python scripts/build_my_pack.py

3. Evaluate — Prove the pack adds value¶

uv run python scripts/eval_single_pack.py my-pack --sample 10

4. Package and Install — Register as an agent skill¶

cd data/packs && tar -czf my-pack.tar.gz my-pack
wikigr pack install my-pack.tar.gz

Packs install to ~/.wikigr/packs/ and auto-register as Claude Code skills. The agent automatically uses the pack when you ask domain questions.

5. Query — The skill activates automatically¶

In Claude Code, just ask. Or use the CLI:

uv run wikigr query "How does X work?" --pack my-pack

For the full walkthrough, see the Tutorial.

Documentation¶

Getting Started¶

Overview — What packs are and when to use them
Quick Start — Build and query your first pack in 5 minutes
Tutorial — Complete walkthrough from domain selection to deployment

Concepts¶

How Packs Work — Content pipeline, query pipeline, retrieval modules
Architecture — System design, data model, technology stack
Retrieval Pipeline — Each stage explained in detail

Evaluation¶

Methodology — How we measure pack quality
Results — Full accuracy data for all 48 packs
Improving Accuracy — Techniques that brought accuracy to 99%

How-To Guides¶

Build a Pack — Step-by-step from URLs to installed skill
Run Evaluations — Single-pack and cross-pack evaluation
Configure the Retrieval Pipeline — Tuning retrieval modules
Handle Exceptions — Catching API, DB, and network errors from WikiGR components

Design¶

Exception Handling — Error domain contracts, removed fallbacks, renamed pipeline stages, security requirements, and known gap inventory

Reference¶

KG Agent API — Constructor, query(), response format
Exception Types — All exception types raised by WikiGR components
Chat API — POST /chat and GET /chat/stream endpoints, SSE event types, examples
Pack Distribution — package_pack / unpackage_pack, archive format, extraction safety
CLI Commands — All wikigr commands
Pack Manifest — manifest.json format
Pack Utilities — load_urls and shared helpers for build scripts
Test Suite — Test files, expected counts, DB-skip guard, mock strategy