Session RAG Eval

Session Attachment RAG Eval

Use this skill when validating file question-answering behavior for large chat attachments.

Goal

Verify both sides of model behavior:

• The model calls query_session_attachment when the answer depends on an uploaded large file.
• The model avoids retrieval for clearly unrelated user requests.

Prefer the Chatbox conversation-flow harness for product validation because it exercises the real renderer, config, license, local API, file upload, indexing, tool registration, and persisted messages.

Fixtures

Default fixture repo:

../../chatbox-session-rag-eval-fixtures

Fixture types:

• synthetic long documents with hidden needle facts;
• Wikipedia-derived real long documents with source attribution;
• direct positive cases;
• implicit positive cases where the user does not explicitly ask to search the file;
• unrelated negative cases.

Regenerate fixtures in the fixture repo:

node scripts/generate-fixtures.mjs
node scripts/fetch-real-fixtures.mjs

Chatbox Flow

Start the local API before running the harness. Then build with USE_LOCAL_API=true; the renderer API origin is compiled into the bundle.

USE_LOCAL_API=true node ./node_modules/electron-vite/bin/electron-vite.js build --mode development
pnpm eval:session-rag:chatbox -- --case long-citrine-threshold --keep-user-data

The harness copies the real config.json into an isolated temporary userDataDir, injects a temporary default chat model if missing, and stores the session RAG sqlite DB at a separate temporary path.

Known Failure Modes

• "This attachment is too large..." means session_attachment_embedding capability is false or unavailable for the active local API/license path.
• If requests go to production, rebuild with USE_LOCAL_API=true.
• If Electron hangs at startup, check stale CDP/Electron processes. The harness now allocates a free CDP port.
• If the assistant says the file is still indexing after DB says ready, inspect the persisted user message. The message file must have sessionAttachmentIndexStatus: "ready" before context building.
• Do not trust a log line that says the message was submitted unless the textarea cleared and the session contains the user message.

Useful Cases

pnpm eval:session-rag:chatbox -- --case long-citrine-threshold
pnpm eval:session-rag:chatbox -- --case implicit-citrine-current-policy
pnpm eval:session-rag:chatbox -- --case real-wiki-apollo-implicit-landing-site
pnpm eval:session-rag:chatbox -- --case multi-turn-real-wiki-apollo-followup
pnpm eval:session-rag:chatbox -- --case unrelated-simple-math
pnpm eval:session-rag:chatbox -- --case real-wiki-unrelated-capital

For fast model-behavior iteration without Electron:

pnpm eval:session-rag -- --dry-run
pnpm eval:session-rag -- --case implicit-citrine-current-policy