Framework Configuration (LlamaIndex and LangChain)¶

LangChain is a full peer framework alongside LlamaIndex. Every pipeline stage — chunking, KG extraction, property graph ingestion/retrieval, vector store, search store, and hybrid fusion — can independently use LangChain via env var pickers.

For architecture details, retriever layers, and source labeling, see LangChain Architecture.

Installation¶

# Core — all 13 LLM providers as native LC clients, Neo4j Cypher retrieval, RDF SPARQL QA fusion,
#        LC vector stores (Qdrant, ES, Chroma, Pinecone, pgvector, LanceDB, Milvus, Weaviate, OpenSearch),
#        LC search (BM25, ES, OpenSearch)
uv pip install -e ".[langchain]"

# Extended graph backends — ArangoDB, Apache AGE, HugeGraph, TigerGraph, Cosmos Gremlin, Spanner
uv pip install --override extras-overrides.txt -e ".[langchain,langchain-extras]"

# SurrealDB — isolated group (dependency conflicts with langchain-extras)
uv pip install -e ".[surrealdb-extras]"

extras-overrides.txt pins requests, urllib3, rich, setuptools, and decorator to prevent hugegraph-python's outdated deps from downgrading the environment.

Framework Pickers¶

Each stage defaults to LlamaIndex and can be switched to LangChain independently:

CHUNKER_BACKEND=llamaindex        # llamaindex | langchain
KG_EXTRACTOR_BACKEND=llamaindex   # llamaindex | langchain
GRAPH_BACKEND=llamaindex          # llamaindex | langchain (auto for LC-only stores)
VECTOR_BACKEND=llamaindex         # llamaindex | langchain
SEARCH_BACKEND=llamaindex         # llamaindex | langchain
RETRIEVAL_FUSION=llamaindex       # llamaindex (QueryFusionRetriever) | langchain (EnsembleRetriever/RRF)

Chunker Backend¶

CHUNKER_BACKEND=langchain
LC_SPLITTER_TYPE=recursive    # recursive | character | token | markdown | python | sentence_transformers

LC chunks are stashed as system._last_lc_chunks and passed directly to LC vector/search stores — no re-embedding.

KG Extraction Backend¶

KG_EXTRACTOR_BACKEND=langchain   # uses LLMGraphTransformer

If KG_EXTRACTOR_BACKEND=langchain and GRAPH_BACKEND=langchain, LC GraphDocument objects are written directly to the LC graph store. If KG_EXTRACTOR_BACKEND=llamaindex and GRAPH_BACKEND=langchain, LI triplets are converted automatically.

Property Graph Backend¶

GRAPH_BACKEND=langchain
PG_GRAPH_DB=arangodb    # or any other supported store — see CONFIG-PROPERTY-GRAPH.md

Property Graph Retriever Toggles¶

TextToGraphQueryRetriever (NL → Cypher/AQL/GSQL/SurrealQL): - Auto-enabled for all LC non-vector stores (ArangoDB, Apache AGE, HugeGraph, SurrealDB, TigerGraph, Cosmos Gremlin, Spanner) — it is their only graph retrieval path. - Auto-enabled for Neo4j LC when LANGCHAIN_PG_VECTOR_SEARCH=false (default). - Suppressed for Neo4j when LANGCHAIN_PG_VECTOR_SEARCH=true (vector + neighborhood takes over). Set USE_LC_TEXT_TO_GRAPH=true to add it back alongside vector retrieval.

USE_LC_TEXT_TO_GRAPH=true   # only needed for Neo4j when LANGCHAIN_PG_VECTOR_SEARCH=true

GraphNeighborhoodRetriever (k-hop graph walk from seed entities; Neo4j only):

USE_PG_NEIGHBORHOOD=true    # default true; auto-enabled when LANGCHAIN_PG_VECTOR_SEARCH=true

GraphEntityVectorRetriever (entity-level vector seeding; Neo4j only; opt-in):

LANGCHAIN_PG_VECTOR_SEARCH=false  # default; set true to enable

LANGCHAIN_PG_VECTOR_SEARCH=true doubles query time (~50s vs ~25s) and auto-enables USE_PG_NEIGHBORHOOD. Only enable if you want entity-level Neo4j vector similarity seeding.

LangChain-Only Property Graph Stores¶

These stores auto-select GRAPH_BACKEND=langchain:

Vector Backend¶

VECTOR_BACKEND=langchain
VECTOR_DB=milvus    # or weaviate, lancedb, chroma, pinecone, postgres, etc.
MILVUS_VECTOR_DB_CONFIG={"host": "localhost", "port": 19530}

Search Backend¶

SEARCH_BACKEND=langchain
SEARCH_DB=elasticsearch
ELASTICSEARCH_SEARCH_DB_CONFIG={"hosts": ["http://localhost:9200"], "index_name": "hybrid_search"}

RDF SPARQL Retrieval¶

RDF_GRAPH_DB=fuseki   # fuseki | graphdb | oxigraph | neptune_rdf | none

When RDF_GRAPH_DB is not none, a LangChain SPARQL QA chain (TextToGraphQueryRetriever → SPARQL) is fused into hybrid search automatically. No USE_LANGCHAIN_RDF flag needed — RDF retrieval is always via LangChain.

Synonym Expansion¶

Expand queries for broader retrieval coverage (opt-in — disabled by default). SYNONYM_EXPLODER_SCOPE is a comma-separated list of retriever tags that limits which retrievers receive the expanded queries:

USE_SYNONYM_EXPLODER=false          # default; set true to enable
SYNONYM_EXPLODER_SCOPE=none         # e.g. llamaindex_search,langchain_search

When enabled, the LLM generates synonyms for query keywords; each synonym is used as an additional embedding query against the scoped retrievers.

Retrieval Fusion¶

RETRIEVAL_FUSION=llamaindex   # QueryFusionRetriever, mode=relative_score (default)
RETRIEVAL_FUSION=langchain    # EnsembleRetriever (RRF) — activates when ALL retrievers are LC-backed

EnsembleRetriever is in langchain_classic.retrievers.ensemble. Falls back silently to QueryFusionRetriever if any LI-native retriever is present.

Source Labels in Results¶

Every search result shows which database it came from. This happens automatically — no configuration needed. See the Search Tab Guide (UI Guide → Tab 3) for examples.