Local RAG Returns Empty Results, Dimension Error (1024 vs 768) or Misses Retrieved Facts

Each embedding model outputs a fixed vector dimension (`all-MiniLM-L6-v2` = 384, `nomic-embed-text` = 768, `bge-m3` = 1024). Existing ChromaDB/pgvector collections remain locked to the dimension of the model used when the first document was uploaded.

1. Lock your embedding model to `bge-m3` (1024-dim) for bilingual ZH/EN corpora.
2. Delete and recreate the knowledge collection (or click "Reset Vector DB & Re-index" in Admin Settings) so all chunks are re-embedded at 1024 dimensions.

# Verify exact output dimension of your active Ollama embedding model:
curl -s http://127.0.0.1:11434/api/embed \
  -d '{"model":"bge-m3","input":"dimension check"}' | python3 -c '
import sys, json
vecs = json.load(sys.stdin).get("embeddings", [[]])
print("Embedding Dimension:", len(vecs[0]))
'

[BOUNDARY]Even if two different models both output 768 dimensions, their vector spaces are incompatible—switching models ALWAYS requires full re-indexing.

Ollama defaults every model to `num_ctx = 2048` unless overridden. When your RAG app prepends 5–8 retrieved chunks (~3,500 tokens) + system instructions, Ollama silently drops the earliest tokens (your top-ranked chunks!) so the LLM never sees them.

1. Set `num_ctx: 8192` (or `16384`) in your chat model parameters or Modelfile.
2. Enable `OLLAMA_FLASH_ATTENTION=1` and `OLLAMA_KV_CACHE_TYPE=q8_0` so the larger context + co-loaded `bge-m3` model stay 100% on GPU.

# Create a RAG-ready 16K context Modelfile
cat << 'EOF' > Modelfile.rag
FROM qwen2.5:14b-instruct-q4_K_M
PARAMETER num_ctx 16384
PARAMETER temperature 0
EOF
ollama create qwen2.5-14b-rag -f Modelfile.rag

[BOUNDARY]Check the CONTEXT column in `ollama ps` during a RAG query to confirm the server actually allocated 8192/16384 tokens.

Standard PDF parsers (`pypdf` / `pdfminer`) only read embedded character streams. Scanned PDFs yield 0 characters, while two-column PDFs merge left and right columns across the same horizontal Y-coordinate.

1. Check if `pdftotext file.pdf - | wc -c` returns < 50 characters per page (indicating a scanned image PDF).
2. Rasterize pages at 150 DPI (`pdftoppm -png -r 150`) and extract clean Markdown tables via local `qwen2.5vl:7b` before indexing into your RAG store.

# Check if PDF has an embedded text layer (if close to 0, it's a scanned image PDF):
pdftotext sample.pdf - | wc -c

# Rasterize at 150 DPI for local Vision LLM OCR:
pdftoppm -png -r 150 sample.pdf ./page

[BOUNDARY]Rasterizing at 150 DPI (~1240×1754 px) is the accuracy/VRAM sweet spot; 300+ DPI quadruples visual tokens without improving OCR accuracy.