bash.sh

#!/bin/bash
#
# Reference requests for both endpoints. This is a cheat-sheet, not a runnable
# suite — copy the command you need. Assumes the server is on :3000.

# --- Ingest a PDF -----------------------------------------------------------

# Linux / macOS
curl -X POST http://localhost:3000/ingest -F "file=@test.pdf"

# Windows PowerShell
curl.exe -X POST http://localhost:3000/ingest -F "file=@test.pdf"

# Response:
# { "message": "Ingested 8 chunks from \"test.pdf\"" }

# --- Ask a question ---------------------------------------------------------

# Linux / macOS
curl -X POST http://localhost:3000/chat \
  -H "Content-Type: application/json" \
  -d '{ "question": "How should I handle errors in async functions?" }'

# Windows PowerShell
curl.exe -X POST http://localhost:3000/chat -H "Content-Type: application/json" -d "{\"question\": \"How should I handle errors in async functions?\"}"

# --- Reading the results ----------------------------------------------------

# A question the document actually answers scores well:
#
#   curl -X POST http://localhost:3000/chat \
#     -H "Content-Type: application/json" \
#     -d '{ "question": "Webhook responded" }'
#
#   {"answer":"Webhook responded with status code 400.",
#    "sources":["test.pdf"],"topSimilarity":"0.694"}
#
# A question it does not answer scores poorly, and the model says so rather
# than inventing an answer — topSimilarity 0.521, and the reply opens with
# "the context provided does not contain information about ...".
#
# Rule of thumb for topSimilarity:
#   > 0.7   retrieval found genuinely relevant chunks
#   < 0.5   retrieval struggled; it matched keywords, not meaning
#
# Worth testing deliberately: ask about something the PDF never mentions. The
# model should decline rather than guess. That is the behaviour you want in
# production, and it comes from the system prompt in src/embeddings.js.