trelix-langchain

LangChain retriever for trelix — semantic code search using Tree-sitter AST parsing, hybrid BM25+vector search, call-graph expansion, and streaming synthesis support.

Install

pip install trelix-langchain

Optional features live on core's extras, not the adapter's — trelix-langchain declares none of its own, so install it alongside the core extra you want.

For AWS Bedrock embeddings (Cohere or Titan):

pip install trelix-langchain "trelix[bedrock]"

For code-optimized embeddings, pick the backend you want:

pip install trelix-langchain "trelix[bge-code]"     # BGE-Code
pip install trelix-langchain "trelix[nomic-code]"   # Nomic-Code
pip install trelix-langchain "trelix[lance]"        # Lance vector backend

With knowledge graph support (NetworkX BFS retrieval leg):

pip install trelix-langchain 'trelix[knowledge-graph]'

Basic Usage

from trelix_langchain import TrelixRetriever

# First index your repo (one-time)
# trelix index /path/to/repo

retriever = TrelixRetriever(repo_path="/path/to/repo", k=10)
docs = retriever.invoke("how does authentication work?")

for doc in docs:
    print(doc.metadata["source"], doc.metadata["score"])
    print(doc.page_content[:200])

Each returned Document carries rich metadata:

Graph-Enhanced Retrieval

Enable the knowledge graph as a 4th retrieval leg for architecture-aware queries:

from trelix_langchain import TrelixRetriever

# Standard hybrid retrieval (v2.1.0: all beast-mode flags default to false)
retriever = TrelixRetriever(repo_path="/path/to/repo", k=10)

# With graph-aware BFS (requires trelix[knowledge-graph])
retriever = TrelixRetriever(
    repo_path="/path/to/repo",
    k=10,
    graph_search_enabled=True,   # enables 4th BFS retrieval leg
    graph_search_depth=2,
)

# v2.1.0: Combine graph search with beast-mode retrieval legs
# (Enable via env vars BEFORE constructing retriever)
retriever = TrelixRetriever(
    repo_path="/path/to/repo",
    k=10,
    graph_search_enabled=True,
)

# Each Document.metadata includes graph source info
docs = retriever.invoke("how does auth relate to the data layer?")
for doc in docs:
    print(doc.metadata["retrieval_source"])  # "graph_search", "file_summary", "vector", "bm25", "pagerank"

When graph_search_enabled=True, the retriever merges results from multiple legs (v2.1.0 adds optional file-summary and PageRank):

Graph BFS surfaces structurally related symbols even when semantic similarity is low — useful for cross-cutting concerns like auth, logging, and rate-limiting that touch many modules.

Graph config options

You can also set these via environment variables (v2.1.0+):

# Enable graph search and all v2.1.0 beast-mode legs
TRELIX_GRAPH_SEARCH_ENABLED=true \
TRELIX_RETRIEVAL_FILE_SUMMARY_LEG=true \
TRELIX_RETRIEVAL_PAGERANK_BOOST=true \
trelix index /path/to/repo

Prerequisite: build the knowledge graph before querying — trelix graph /path/to/repo. The graph is persisted in <repo>/.trelix/ and reused across retriever calls.

LangChain RAG Chain (LCEL)

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
from trelix_langchain import TrelixRetriever

retriever = TrelixRetriever(repo_path="/path/to/repo", k=8)

prompt = ChatPromptTemplate.from_template(
    "Answer the question using only the code context below.\n\n"
    "Context:\n{context}\n\n"
    "Question: {question}"
)

def format_docs(docs):
    return "\n\n".join(
        f"# {d.metadata['source']} ({d.metadata['symbol']})\n{d.page_content}"
        for d in docs
    )

chain = (
    {"context": retriever | format_docs, "question": RunnablePassthrough()}
    | prompt
    | ChatOpenAI(model="gpt-4o")
    | StrOutputParser()
)

answer = chain.invoke("How does the authentication middleware work?")
print(answer)

RetrievalQA (classic interface)

from langchain.chains import RetrievalQA
from langchain_openai import ChatOpenAI
from trelix_langchain import TrelixRetriever

retriever = TrelixRetriever(repo_path="/path/to/repo", k=10)
llm = ChatOpenAI(model="gpt-4o")

qa = RetrievalQA.from_chain_type(
    llm=llm,
    retriever=retriever,
    return_source_documents=True,
)

result = qa.invoke({"query": "Where is rate limiting applied?"})
print(result["result"])
for doc in result["source_documents"]:
    print(" -", doc.metadata["source"])

Configuration

You can also set the provider directly on the retriever instance:

retriever = TrelixRetriever(repo_path="/path/to/repo", provider="openai", k=10)

Provider Switching (v2.0.0+, updated v2.4.0)

# bge-code is EXPERIMENTAL: pooling unverified, no quality claim (see trelix docs/PROVIDERS.md)
TRELIX_EMBEDDER_PROVIDER=bge-code trelix index /path/to/repo

# Use Nomic-Code embeddings
TRELIX_EMBEDDER_PROVIDER=nomic-code trelix index /path/to/repo

# Use Bedrock Cohere embeddings (reuses AWS credentials)
TRELIX_EMBEDDER_PROVIDER=bedrock-cohere trelix index /path/to/repo

# Use Azure OpenAI embeddings
TRELIX_EMBEDDER_PROVIDER=azure trelix index /path/to/repo

# Use local sentence-transformers (no API key needed, works offline)
TRELIX_EMBEDDER_PROVIDER=local trelix index /path/to/repo

The index and the retriever must use the same provider — re-index whenever you switch.

Beast-Mode Retrieval (v2.1.0+)

trelix v2.1.0 adds five opt-in retrieval improvements — HyDE (hypothetical document expansion), FLARE (active retrieval), file-summary leg, PageRank boost, and telemetry — all activated via environment variables. No code changes needed:

from trelix_langchain import TrelixRetriever

# v2.1.0: Enable beast-mode features via env vars before constructing retriever
# Export any or all of these (all default to false):
# TRELIX_RETRIEVAL_HYDE_FALLBACK=true        # HyDE: expand queries with hypothetical docs
# TRELIX_RETRIEVAL_FILE_SUMMARY_LEG=true    # Add file-summary retrieval leg
# TRELIX_RETRIEVAL_PAGERANK_BOOST=true      # Boost symbols by PageRank centrality
# TRELIX_TELEMETRY_ENABLED=true             # Record per-query telemetry

retriever = TrelixRetriever(
    repo_path="/path/to/repo",
    provider="azure",  # or "local", "openai"
    k=10,
)
docs = retriever.invoke("how does the authentication system work?")

What's New in v2.1.0:

• HyDE fallback: If semantic search scores are low, generate hypothetical docs and re-score
• File-summary leg: Index-time file summaries as a 5th retrieval source (cross-file context)
• PageRank boost: Upweight symbols in call-graph "hub" positions
• Telemetry: Opt-in metrics on retrieval latency, source distribution, and cache hit rates
• All features are zero-overhead when off — use env vars to opt in per deployment

Streaming Synthesis (v2.0.0+)

Streaming synthesis support for real-time code context generation:

from trelix_langchain import TrelixRetriever, StreamingSynthesizer
from langchain_openai import ChatOpenAI

retriever = TrelixRetriever(repo_path="/path/to/repo", k=8)
synthesizer = StreamingSynthesizer(
    llm=ChatOpenAI(model="gpt-4o"),
    retriever=retriever
)

# Streamed synthesis output
for chunk in synthesizer.synthesize_stream("How does the auth flow work?"):
    print(chunk, end="", flush=True)

GitHub PR Review (v2.4.0+)

Fetch a pull request diff from GitHub and run DiffReviewer directly through the retriever:

from trelix_langchain import TrelixRetriever

retriever = TrelixRetriever(repo_path="/path/to/repo", k=8)

# Retrieve context relevant to a PR diff
# Use the trelix CLI: trelix review --pr owner/repo#42
# Or post review comments: trelix review --pr owner/repo#42 --post-comments
# Requires GITHUB_TOKEN env var

Set GITHUB_TOKEN in your environment. The integration fetches all changed files in the PR, retrieves relevant code context for each diff hunk, and can optionally post a single batched review back to GitHub.

MCP Pagination (v2.4.0+)

The search_code MCP tool now returns a pagination envelope instead of a raw list. If you call trelix-mcp from LangChain tool wrappers, update your iteration:

# v2.4.0+ response shape from search_code MCP tool
response = search_code_tool.run({"query": "auth", "repo_path": "/repo"})
# response = {"results": [...], "next_cursor": 10, "total_available": 25}

for result in response["results"]:
    print(result)

# Paginate: pass next_cursor as cursor= in the next call

Multi-Query Expansion Observability (v2.4.0+)

When multi_query_enabled=True in your IndexConfig, the retriever now surfaces expansion telemetry via the ExpandResult dataclass:

from trelix_langchain import TrelixRetriever
from trelix.retrieval import MultiQueryExpander

expander = MultiQueryExpander(llm=your_llm)
expand_result = expander.expand("how does auth work?")
# expand_result.queries       — list of sub-queries generated
# expand_result.llm_used      — model name
# expand_result.elapsed_ms    — wall-clock time for expansion

Expansion metadata (expansion_used, expansion_variants, expansion_elapsed_ms) is persisted automatically to the query_telemetry table. Existing databases are upgraded automatically via an idempotent ALTER TABLE ADD COLUMN migration.

FederatedRetriever Cache (v2.4.0+)

When using FederatedRetriever across multiple repos, enable the TTL cache to avoid redundant retrievals within a debugging session:

from trelix.retrieval import FederatedRetriever

retriever = FederatedRetriever(registry=my_registry, cache_ttl=120.0)
# cache_ttl=0 disables caching
stats = retriever.cache_stats()   # {"hits": 42, "misses": 5, "size": 18}
retriever.clear_cache()           # force eviction

The cache is SHA-256-keyed, thread-safe, and scoped to the process lifetime. Expected ~90% hit rate for typical debugging-session query patterns.

Multi-Repo Watching (v2.4.0+)

Watch multiple repos simultaneously and keep their indexes live:

# CLI
trelix watch-all

# Watches all registered repos; shows per-repo stats on exit; Ctrl+C to stop

from trelix.watchers import MultiRepoWatcher

watcher = MultiRepoWatcher(repo_paths=["/repo/a", "/repo/b"])
await watcher.watch()  # uses watchfiles under the hood; hash guard prevents cascade re-index

Configuration (v2.4.0+)

In addition to the env vars above, v2.4.0 adds:

TRELIX_RETRIEVAL_FLARE_MAX_ITER is still accepted but emits a DeprecationWarning. It will be removed in v3.0.0.

Links

• trelix on GitHub
• trelix on PyPI
• trelix-mcp — MCP server for Claude Code, Cursor, Windsurf
• trelix-llama-index — LlamaIndex retriever