chunksmith-agent 0.4.0

ChunkSmith document Q&A agent over saved multi-indexing outlines.

chunksmith-agent

Standalone document Q&A over saved ChunkSmith index JSON (no dependency on chunksmith-core, chunksmith-multimodal, or chunksmith-pageindex).

Install

pip install chunksmith-agent
pip install "chunksmith-agent[langchain]"   # LangChain tool-calling Q&A

Usage

from pathlib import Path
from chunksmith_agent import ChunkSmithAgent
from chunksmith_agent.index_builder import build_document_index_from_saved

index = build_document_index_from_saved(
    pageindex_path=Path("runs/my-doc/json/my-doc_pageindex.json"),
)
agent = ChunkSmithAgent(index)
answer = agent.ask("What is this document about?")
print(answer.answer)

JSON artifact contract

The agent reads files produced by chunksmith-cli (or compatible tools) under a run folder:

File	Fields used
json/*_pageindex.json	doc_name, structure, optional embedded canonical_bundle
json/*_canonical_bundle.json	elements[], coded_formate, path_image
Outline nodes	node_id, title, summary, start_index/end_index, anchor fields

Environment variables

Same LLM env vars as ChunkSmith CLI / MVL:

• OPENAI_API_KEY (or CHATGPT_API_KEY)
• PAGEINDEX_MODEL, CHUNKSMITH_LLM_MODEL, LLM_MODEL
• Azure: AZURE_API_KEY, AZURE_API_BASE, AZURE_API_VERSION

Integration patterns (loose coupling)

Do not depend on chunksmith-adapters inside this package. Pass data in yourself:

Caller	How to load index	Agent config
CLI	build_document_index_from_saved(pageindex_path=...)	load_settings() from .env
MVL app	chunksmith_agent_bridge.load_document_index_from_mvl(repo, ...)	load_settings() or explicit AgentSettings
Custom app	Fetch JSON from your DB/S3 → build_document_index(dict)	Your env / settings

Install separately from CLI or MVL:

pip install "chunksmith-agent[langchain]"

Extensibility

• Outline nodes: come from saved JSON (structure); re-index or edit JSON to add sections.
• Tools: extend _make_tools() in tool_agent.py (e.g. add get_page_images).
• Another agent in your app: compose ChunkSmithAgent alongside your own planners/retrievers — this package is one document Q&A brain, not your whole system.