103 lines
3.4 KiB
Markdown
103 lines
3.4 KiB
Markdown
# Reranker sidecar — llama.cpp + jina-reranker-v2-base
|
||
|
||
Phase 6 setup. The MCP server reads `RERANK_URL` and, when set, pipes
|
||
the top-50 dense (or hybrid) chunks through this sidecar before
|
||
returning to the LLM. See `docs_mcp/server.py:_rerank_pool`.
|
||
|
||
## Production deploy — trashpanda (Tesla P4, 8 GB VRAM)
|
||
|
||
This is where the reranker lives. Same box that runs the Drawbar
|
||
backend + Cloudflare Tunnel, so the MCP server can reach it on the
|
||
internal LAN.
|
||
|
||
```bash
|
||
ssh [email protected] \
|
||
'docker run -d --name llama-rerank --restart unless-stopped --gpus all \
|
||
-p 8082:8080 \
|
||
ghcr.io/ggml-org/llama.cpp:server-cuda \
|
||
-hf gpustack/jina-reranker-v2-base-multilingual-GGUF:Q8_0 \
|
||
--reranking --host 0.0.0.0 --port 8080 -ngl 99'
|
||
```
|
||
|
||
Key flags:
|
||
- `--gpus all` — pass through the Tesla P4
|
||
- `server-cuda` image — CUDA-built llama.cpp (not the CPU-only `:server`)
|
||
- `-ngl 99` — offload all layers to GPU
|
||
- `-hf <repo>` — auto-download from HuggingFace on first start (~280 MB,
|
||
cached in the container volume)
|
||
- `--reranking` — enables `/v1/rerank` endpoint
|
||
- `--restart unless-stopped` — survives reboot
|
||
|
||
VRAM usage: ~280 MB model + CUDA context. Well under the 8 GB the
|
||
Tesla P4 has, leaves room for nomic-embed-text (~560 MB) if you
|
||
later co-host it.
|
||
|
||
## Configure the MCP server
|
||
|
||
In production the MCP reaches this sidecar **over the Drawbar compose
|
||
network by service name**, not by host IP — `RERANK_URL` is baked into
|
||
the image as:
|
||
|
||
```
|
||
RERANK_URL=http://llama-rerank:8080
|
||
```
|
||
|
||
so the deployed `chem-mcp` service sets nothing. See
|
||
`deploy/docker-compose.yml`.
|
||
|
||
That only resolves if the `llama-rerank` container is attached to
|
||
`drawbar-backend_default`. If it is on the default bridge network
|
||
instead, the name resolves via public DNS to an unrelated IP and
|
||
connection-refuses — and `search_docs` falls back to dense+BM25
|
||
**silently**. Attach it with:
|
||
|
||
```bash
|
||
docker network connect drawbar-backend_default llama-rerank
|
||
```
|
||
|
||
For local dev outside that network, point at the published port
|
||
directly:
|
||
|
||
```bash
|
||
export RERANK_URL=http://10.10.1.65:8082
|
||
```
|
||
|
||
## Verify
|
||
|
||
```bash
|
||
curl http://10.10.1.65:8082/v1/rerank -H 'Content-Type: application/json' -d '{
|
||
"query": "soybean herbicide for waterhemp",
|
||
"documents": [
|
||
"Roundup Custom for fallow burndown",
|
||
"Sencor metribuzin controls waterhemp in soybean pre-emergence"
|
||
]
|
||
}'
|
||
```
|
||
|
||
Expect index=1 (the Sencor doc) at score ~0.8, index=0 at a strongly
|
||
negative score, in under 1 s.
|
||
|
||
## Performance reference
|
||
|
||
| Mode | Pool | Wall time |
|
||
|---|---|---|
|
||
| CPU (local 28-thread Xeon) | 50 docs | ~23 s |
|
||
| GPU (Tesla P4 on trashpanda) | 50 docs | ~0.7-1.5 s |
|
||
| GPU (Tesla P4) | 20 docs | ~0.4 s |
|
||
|
||
The Tesla P4 is Pascal-era (8.1 TFLOPs FP32) so a modern Ampere or
|
||
Ada Lovelace GPU would be ~3-5× faster, but for the row-crop label
|
||
corpus query rate the P4 is plenty.
|
||
|
||
## Troubleshooting
|
||
|
||
- **Model not on GPU?** Check `docker logs llama-rerank | grep CUDA` —
|
||
you should see `CUDA0 : Tesla P4 (8109 MiB, ... free)` and tensor
|
||
load lines. If you see CPU-only init, you forgot `--gpus all` or
|
||
used `:server` instead of `:server-cuda`.
|
||
- **Conflict with Ollama on the same GPU?** No — both processes can
|
||
share the GPU, CUDA handles VRAM partitioning. nomic-embed-text +
|
||
jina-reranker-v2-base together use ~840 MB on the 8 GB card.
|
||
- **First rerank call is slow (~4 s)?** Warm-up. Subsequent calls are
|
||
~0.7 s for 50 docs.
|