feat: add parish-rag demo — RAG-enhanced NPC knowledge

[dmooney]

Summary

Standalone parish-rag crate demonstrating how retrieval-augmented generation can ground NPC dialogue in the parish's own lore (world locations, NPC bios, festivals) instead of relying on the model's training data or the 4 hand-written knowledge bullets in npcs.json.

The crate is a pattern demo, not a replacement for the existing keyword recall in parish-npc::memory. No existing code was modified.

What's in it

• LoreIndex — in-memory vector store with cosine top-k search.
• AnyEmbedder — unified handle over two backends:
  • HashEmbedder — deterministic hashing-trick embedder. Zero network, zero dependencies, reproducible test output.
  • OllamaEmbedder — real /api/embeddings client (e.g. nomic-embed-text) for semantic retrieval.
• build_rundale_corpus(path) — chunks mods/rundale/{world,npcs,festivals}.json into ~280 single-fact passages.
• format_recall_block(hits) — formats retrieved hits as a "KNOWLEDGE YOU RECALL" block to append to an NPC's system prompt.
• npc_knowledge_demo binary — runs a scripted set of questions, printing retrieved passages + baseline vs. RAG system prompts. With --llm, calls an OpenAI-compatible chat endpoint for both and prints the responses side by side.

Try it

# Offline — works anywhere, no Ollama needed
cargo run -p parish-rag --bin npc_knowledge_demo

# Live — semantic embeddings + real LLM responses
cargo run -p parish-rag --bin npc_knowledge_demo -- \
    --embedder ollama --embed-model nomic-embed-text \
    --chat-model qwen2.5:7b --llm

# Another NPC / custom question
cargo run -p parish-rag --bin npc_knowledge_demo -- \
    --npc "Siobhan Murphy" \
    --question "Who should I see about renting farmland?"

Sample offline output (Padraig, top-4):

Q: Tell me about the crossroads — is there anything strange about it?
  [0.250] location:The Crossroads — folklore: Crossroads hold power in Irish folklore — a place between places, where the veil is thin.
  [0.250] npc:Siobhan Murphy — Siobhan Murphy is a Farmer (45 years old).
  ...
Q: What is Lughnasa and when is it celebrated?
  [0.213] festival:Lughnasa — Lughnasa is a festival held on the 1st of August. Start of autumn — the harvest festival
  ...

Test plan

• cargo test -p parish-rag — 31 tests, all deterministic, offline, pass.
• cargo clippy -p parish-rag --all-targets -- -D warnings — clean.
• cargo fmt -p parish-rag — formatted.
• cargo check --workspace --exclude parish-tauri — clean. (Tauri excluded only because of a pre-existing gdk-pixbuf-sys system-library issue in this sandbox, unrelated to this change.)
• cargo run -p parish-rag --bin npc_knowledge_demo — demo runs end-to-end against shipped Rundale mod.
• Live mode with Ollama — needs a machine with Ollama + nomic-embed-text installed.

Files changed

• Cargo.toml — added crates/parish-rag to workspace members.
• Cargo.lock — regenerated for the new crate.
• crates/parish-rag/ — new crate (lib + bin + README + tests).

No changes to parish-core, parish-npc, parish-inference, or any existing code.

https://claude.ai/code/session_01Npr5LoSFHrwKoYsQJVUzd7

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c1c6f591f2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Normalize chat base URL before appending the v1 route

Constructing the chat URL with format!("{}/v1/chat/completions", base_url.trim_end_matches('/')) breaks whenever callers pass a common OpenAI-style base URL that already ends in /v1 (for example http://localhost:11434/v1), producing /v1/v1/chat/completions and consistent 404s. This makes --llm fail for a realistic configuration; stripping a trailing /v1 before concatenation (as done elsewhere in the repo) avoids that regression.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Return no hits when retrieval scores have no signal

The search path always truncates to k documents even when every cosine score is 0.0, which happens for empty/stopword-only hash queries; this injects arbitrary lore into the prompt despite no semantic match. In practice, queries like "the and of" return the first indexed docs with zero scores, degrading answer quality and making RAG output misleading. Filtering out non-positive scores (or short-circuiting when query magnitude is zero) would prevent false recalls.

Useful? React with 👍 / 👎.