Migration

This page discusses migration from LangChain GraphVectorStore as well as between versions of graph-retriever and langchain-graph-retriever.

From LangChain GraphVectorStore

LangChain GraphVectorStore relied on putting specially crafted Link instances into metadata["links"]. Many cases used link extractors to compute these links, but it was also often useful (and necessary) to create them manually.

When converting from a GraphVectorStore to the new langchain-graph-retriever library, you need to do the following:

1. Replace uses of the link extractors with document transformers.
2. Replace manualy link creation with metadata fields.
3. Replace GraphVectorStore usage with the GraphRetriever.

Replace Link Extractors with Document Transformers

GLiNERLinkExtractor

Replace with GLiNERTransformer, which will populate metadata fields for each label.

HierarchyLinkExtractor

If you already have a parent ID in the metadata, you can remove this. Otherwise, replace with the ParentTransformer which populates a parent field computed from a path. The parent field may be used with edges to achieve parent-to-child, child-to-parent, and sibling-to-sibling navigation.

HtmlLinkExtractor

Replace with HyperlinkTransformer which extracts hyperlinks from each chunk and writes them to a metadata field.

KeybertLinkExtractor

Replace with KeybertTransformer, which will populate a metadata field with the keywords.

Replace Manual Link Creation with Metadata Fields

With the old library, you had to choose the direction of the links when they were created -- either in, out or bidir. With the new library, you simply create the corresponding fields and choose the direction of edges when you issue a query (see next section).

GraphVectorStore Links (Old)

# Document metadata for a page at `http://somesite` linking to some other URLs
# and a few keyword links.
doc = Document(
    ...,
    metadata = {
        "links": [
            Link.incoming("url", "http://somesite"),
            Link.outgoing("url", "http://someothersite"),
            Link.outgoing("url", "http://yetanothersite"),
            Link.bidir("keyword", "sites"),
            Link.bidir("keyword", "web"),
        ]
    }
)

LangChain Graph Retriever (New)

doc = Document(
    ...,
    metadata = {
        "url": "http://somesite",
        "hrefs": ["http://someothersite", "http://yetanothersite"],
        "keywords": ["sites", "web"],
    }
)

These metadata fields can be used to accomplish a variety of graph traversals. For example:

• edges = [("hrefs", "url"), ...] navigates from a site to the pages it links to (from hrefs to url).
• edges = [("keywords", "keywords"), ...] navigates from a site to other sites with the same keyword.
• edges = [("url", "hrefs"), ...] navigates from a site to other sites that link to it.

Per-Query Edges

You can use different edges for each query, allowing you to navigate different directions depending on the needs. In the old library, you only ever navigated out from a site to the things it linked to, while with the new library the metadata captures the information (what URL is this document from, what URLs does it reference) and the edges determine which fileds are traversed at retrieval time.

Replace GraphVectorStore with the GraphRetriever

Finally, rather than creating the links and writing them to a GraphVectorStore you write the documents (with metadata) to a standard VectorStore and apply a GraphRetriever:

LangChain Graph Retriever (New)

from langchain_graph_retriever import GraphRetriever
retriever = GraphRetriever(
    store=vector_store,
    edges=[("hrefs", "url"), ("keywords", "keywords")],
)