[feat] Add IVF+PQ FAISS flag for compressed-memory deployments (air-gapped / Docker-free parity with Qdrant)

Motivation

The FAISS backend today uses IndexIVFFlat, which stores every vector at full float32 precision. For the default embedding model (Qwen/Qwen3-VL-Embedding-2B, dim = 2048), a 30 M-vector corpus requires roughly 245 GB of RAM just to hold the index — far outside the reach of air-gapped or Docker-free deployments that cannot run a Qdrant service.

FAISS ships IndexIVFPQ (Product Quantization) out of the box. Switching to PQ with M=64, nbits=8 compresses each vector from 8 192 bytes to 64 bytes (128× reduction, ~1.9 GB for 30 M vectors) at a typical recall-at-10 drop of ≤ 5–8% on inner-product tasks — a trade-off Qdrant users can already make today via --qdrant-quantization-config. Adding an equivalent flag to the FAISS path closes that gap for users who cannot run a Docker sidecar.

Requested change

Add opt-in --pq-m / --pq-nbits flags to the faiss build path so users can select IndexIVFPQ instead of IndexIVFFlat.

CLI — embed/src/pixelrag_embed/index.py

 # inside the 'build' subparser, faiss-only flags
+p_build.add_argument(
+    "--pq-m",
+    type=int,
+    default=0,
+    help="PQ sub-quantizers (0 = IVFFlat, no compression). "
+         "Must divide the embedding dim evenly. "
+         "Typical: dim/32 — e.g. 64 for dim 2048.",
+)
+p_build.add_argument(
+    "--pq-nbits",
+    type=int,
+    default=8,
+    choices=[4, 8],
+    help="Bits per PQ sub-quantizer (default: 8).",
+)

In build_ivf(), when pq_m > 0, replace:

index = faiss.IndexIVFFlat(quantizer, dim, nlist, metric_type)

with:

index = faiss.IndexIVFPQ(quantizer, dim, nlist, pq_m, pq_nbits)

Config — pixelrag.yaml

index:
  backend: faiss
  pq_m: 64        # sub-quantizers; must divide dim evenly
  pq_nbits: 8     # 4 or 8 (default 8)

Pipeline forwarding — index/src/pixelrag_index/pipelines.py

The faiss branch of the cmd build in build() already passes --nlist. It would additionally pass --pq-m / --pq-nbits when they appear in index_cfg.

Why not just use Qdrant?

Qdrant scalar/binary quantization via --qdrant-quantization-config is an excellent solution for connected deployments. IVF+PQ covers the complementary case: no Docker available, no network service permitted, fully air-gapped lab or edge environments. Both users deserve a compression path; this closes the gap on the FAISS side.

Memory impact at a glance

Figures are illustrative for inner-product search on visual-embedding benchmarks; actual recall depends on the corpus. Even the most aggressive setting (M=64, nbits=4) makes a 30 M-vector corpus fit on a workstation with 4 GB of free RAM.

Scope

• embed/src/pixelrag_embed/index.py: build_ivf() function + arg parser — approximately 25 lines changed/added.
• index/src/pixelrag_index/pipelines.py: forward pq_m / pq_nbits from index_cfg — approximately 6 lines.
• No new dependencies: faiss-cpu and faiss-gpu already include IndexIVFPQ.

Happy to open a PR if this direction looks good to the maintainers.