Service Tei Reranker

5.2.52. TEI Reranker

Image: ghcr.io/huggingface/text-embeddings-inference — CPU: cpu-1.9 (amd64) / cpu-arm64-latest (arm64), GPU: :1.9 Container port: 80 · Default host port: allocated by topology.py slot allocator (LLM band 63040–63049) Default: disabled

1. Overview

Hugging Face text-embeddings-inference (TEI) running mixedbread-ai/mxbai-rerank-base-v1, a cross-encoder that scores (query, passage) pairs. Use it to reorder results from any first-stage retriever (vector, BM25 or hybrid). The image exposes /rerank and a /health probe.

Why this model: mxbai-rerank-base-v1 ships ONNX weights, which the amd64 ORT backend in cpu-1.9 needs. It is also small enough for the arm64 Candle backend in cpu-arm64-latest to finish warmup on Apple Silicon.

Any consumer that sends TEI's body shape (query plus texts) can call it. LightRAG's built-in Jina/Cohere clients send query plus documents, which TEI rejects, so Atlas never wires LightRAG directly to TEI. LightRAG uses the backend rerank adapter (POST /lightrag/rerank), which translates between the two shapes. Enable it with LIGHTRAG_RERANK_ADAPTER_ENABLED=true (see the backend README §5.1).

2. Source variants

Source	Container scale	Endpoint	Notes
container-cpu	1	http://tei-reranker:80	Default CPU image; runs on any host
container-gpu	1	http://tei-reranker:80	CUDA image. Compose does not request a GPU yet (#1373), so this variant gets no GPU access.
localhost	0	http://host.docker.internal:${TEI_RERANKER_LOCALHOST_PORT}	Host-installed TEI
disabled	0	""	Reranker service off

For every enabled source, Kong generates the rerank.localhost alias. Container variants route it to http://tei-reranker:80/; localhost routes it to host.docker.internal:${TEI_RERANKER_LOCALHOST_PORT}.

Container callers can use http://tei-reranker:80. Host callers use the published http://localhost:${TEI_RERANKER_PORT} for a container variant, or the host TEI port for localhost.

3. Configuration

TEI_RERANKER_SOURCE=disabled                       # default
TEI_RERANKER_PORT=...                              # slot-allocated
TEI_RERANKER_LOCALHOST_PORT=63049                  # host-installed TEI rerank port
TEI_RERANKER_MODEL_ID=mixedbread-ai/mxbai-rerank-base-v1
TEI_RERANKER_REVISION=800f24c113213a187e65bde9db00c15a2bb12738
TEI_RERANKER_MAX_CLIENT_BATCH_SIZE=32
TEI_RERANKER_MEMORY_LIMIT=4g
TEI_RERANKER_CPU_LIMIT=2.0
TEI_RERANKER_HF_CACHE_DIR=/data

4. Usage

Run these Bash examples from the Atlas repository root with uv and the bootstrapper dependencies available. Use a running endpoint for the selected source. TEI_RERANKER_PORT defaults to 63041.

Each example reads the active configuration file through Atlas's configuration reader. The default file is .env; set ATLAS_ENV_FILE to select another file. Missing files, invalid assignments, and blank required values stop the example before its request.

# Rerank passages
TEI_RERANKER_PORT=$(PYTHONPATH=bootstrapper uv run --project bootstrapper python - <<'PYCONFIG'
from core.config_parser import ConfigParser

config = ConfigParser(root_dir=".")
if not config.env_file_path.is_file():
    raise SystemExit("Active configuration file is missing or is not a file")
value = config.parse_env_file().get("TEI_RERANKER_PORT", "")
if not value.strip():
    raise SystemExit("Set TEI_RERANKER_PORT in the active configuration file")
print(value, end="")
PYCONFIG
) || exit 1
curl -s "http://localhost:${TEI_RERANKER_PORT}/rerank" \
  -H 'Content-Type: application/json' \
  -d '{
    "query": "What is graph-augmented RAG?",
    "texts": [
      "LightRAG combines knowledge graphs with dense vector retrieval.",
      "GraphQL is a query language.",
      "Reranking improves RAG quality by ordering retrieved passages."
    ]
  }'
# → [{"index": 0, "score": ...}, ...]

4.1. Stack-standard rerank via LiteLLM

When TEI_RERANKER_SOURCE is not disabled and the endpoint resolves, litellm-init registers a tei-rerank model on LiteLLM. Consumers then get a Cohere-shaped POST /v1/rerank in front of TEI, with LiteLLM's auth, cost logging and retries:

LITELLM_PORT=$(PYTHONPATH=bootstrapper uv run --project bootstrapper python - <<'PYCONFIG'
from core.config_parser import ConfigParser

config = ConfigParser(root_dir=".")
if not config.env_file_path.is_file():
    raise SystemExit("Active configuration file is missing or is not a file")
value = config.parse_env_file().get("LITELLM_PORT", "")
if not value.strip():
    raise SystemExit("Set LITELLM_PORT in the active configuration file")
print(value, end="")
PYCONFIG
) || exit 1
LITELLM_MASTER_KEY=$(PYTHONPATH=bootstrapper uv run --project bootstrapper python - <<'PYCONFIG'
from core.config_parser import ConfigParser

config = ConfigParser(root_dir=".")
if not config.env_file_path.is_file():
    raise SystemExit("Active configuration file is missing or is not a file")
value = config.parse_env_file().get("LITELLM_MASTER_KEY", "")
if not value.strip():
    raise SystemExit("Set LITELLM_MASTER_KEY in the active configuration file")
print(value, end="")
PYCONFIG
) || exit 1
curl -s "http://localhost:${LITELLM_PORT}/v1/rerank" \
  -H "Authorization: Bearer ${LITELLM_MASTER_KEY}" \
  -H 'Content-Type: application/json' \
  -d '{"model":"tei-rerank","query":"…","documents":["…","…"]}'
# → {"results": [{"index": 0, "relevance_score": ...}, ...]}

• Provider prefix. /rerank is the Cohere-shaped API ({query, documents}), not an OpenAI modality. LiteLLM registers TEI through the huggingface/ rerank provider, which translates the request into TEI's {query, texts}. The infinity, jina and cohere prefixes send {query, documents}, which TEI rejects, so huggingface/ is pinned.
• Backend adapter vs LiteLLM. The backend /lightrag/rerank adapter serves LightRAG's client shape. LiteLLM /v1/rerank is the standard path for other consumers. No TEI API key is needed: TEI is unauthenticated in-network, and init writes the endpoint into config.yaml.

5. Capabilities & limitations

Support tier: experimental — Capability contract declared; no cited cold-start, workflow, or upgrade qualification run yet (evidence at v0.1.0).

Capability	Status	Verification	Notes
Cross-encoder reranking sources	supported	tested	Atlas resolves architecture-specific amd64 ORT and arm64 Candle CPU images, an NVIDIA image, or an existing host TEI endpoint for the configured model.
LiteLLM standard rerank route	supported	tested	The tei-rerank alias uses LiteLLM's Hugging Face adapter to translate Cohere-shaped documents into TEI text pairs behind gateway authentication.
Direct LightRAG-to-TEI reranking	not-supported	tested	LightRAG sends a documents payload that native TEI rejects; it must use the opt-in backend adapter rather than the TEI endpoint directly.
Authenticated native reranker access	partial	tested	The LiteLLM route requires its master key, but TEI's host-published port and CORS-only rerank.localhost alias expose the unauthenticated native API.
Arbitrary reranker model portability	partial	tested	The default model revision is pinned to the tested 800f24c113213a187e65bde9db00c15a2bb12738 commit for reproducible ONNX amd64 and safetensors arm64 artifacts. Operator model or revision overrides are not pre-certified for both backends or their memory limits.

6. Health checks

Run this Bash example from the Atlas repository root with uv and the bootstrapper dependencies available.

TEI_RERANKER_PORT=$(PYTHONPATH=bootstrapper uv run --project bootstrapper python - <<'PYCONFIG'
from core.config_parser import ConfigParser

config = ConfigParser(root_dir=".")
if not config.env_file_path.is_file():
    raise SystemExit("Active configuration file is missing or is not a file")
value = config.parse_env_file().get("TEI_RERANKER_PORT", "")
if not value.strip():
    raise SystemExit("Set TEI_RERANKER_PORT in the active configuration file")
print(value, end="")
PYCONFIG
) || exit 1
curl -fs "http://localhost:${TEI_RERANKER_PORT}/health"   # 200 OK when up

Container start_period is 300 s, because the first run downloads the model before the server binds.

7. Troubleshooting

• First boot logs optional HuggingFace artifact 404s — expected for some reranker models. TEI probes optional Sentence Transformers files, logs 404 warnings when they are absent, then continues with the model artifacts it needs.
• Out of memory on CPU variant — raise TEI_RERANKER_MEMORY_LIMIT (default 4g) or lower TEI_RERANKER_MAX_CLIENT_BATCH_SIZE.
• Slow inference — CPU is the only working container variant until GPU device requests are wired (#1373). Lower TEI_RERANKER_MAX_CLIENT_BATCH_SIZE, or use the localhost source with a host TEI that has a GPU.
• Model not found — verify TEI_RERANKER_MODEL_ID matches a public HF repo. Private repos need an HF_TOKEN env var (not wired by default; hand-add to the compose env block).

8. Dependencies & Integrations

8.1. Current — Upstream (this service calls)

No upstream calls.

8.2. Current — Downstream (services that call this)

Service	Category
kong	infra
litellm	llm
backend	apps

8.3. Architecture diagram

Open the full-size diagram for a full-screen view.

8.4. Future — Missing pair integrations

No high-confidence opportunities identified.

8.5. Future — Candidate new services

No high-confidence opportunities identified.

8.6. Future — Unused features in this service

No high-confidence opportunities identified.