ConstructionBot

Procore RFI dashboard for the gal tenant. Phase 1 (Layer A): operator drafts RFIs, attaches vendor PDFs, submits to the Procore sandbox. Phase 3 (Layer B): sage-LLM extraction from uploaded RFI PDFs with a human review gate before any Procore write.

Architecture

• Rust/axum + askama + htmx — single binary, tenant Postgres, tokio background tasks in-process.
• No secrets in the repo — DATABASE_URL, Procore client credentials, and SAGE_API_KEY arrive via env vars (manifest dit.secrets, resolved on the runner from OpenBao). Non-secret config in config.toml.
• Human gate — every RFI passes through operator review before Procore. The extraction pre-fills the form; the operator edits and submits.
• sha256-keyed cache — same PDF → same cached extraction, zero double sage compute (phase3-draft AC-3).

Quick start (local dev)

export DATABASE_URL=postgres://constructionbot:pw@localhost:5432/constructionbot
export PROCORE_CLIENT_ID=...     # sandbox DMSA client id
export PROCORE_CLIENT_SECRET=... # sandbox DMSA client secret
export SAGE_API_KEY=...          # sage auth (B0 confirms requirement)

cargo run -- --check-config   # validate config.toml
cargo run                     # serve on :8080

Tests

cargo test --lib               # unit tests (offline, no DB needed)
cargo clippy --all-targets -- -D warnings   # lint (clean)
scripts/no_hardcoded_endpoints.sh           # endpoint gate

# Integration tests (need DB + pdftotext):
nix-shell -p poppler-utils --run \
  'DATABASE_URL=postgres://... cargo test --test extraction_cache -- --ignored --nocapture'

# Corpus worksheet (need DB + pdftotext + sage):
nix-shell -p poppler-utils --run \
  'DATABASE_URL=... SAGE_API_KEY=... cargo test --test extraction_corpus -- --ignored --nocapture'

Phase 3 calibration — pilot week

Procedure

After deploying Layer B, run the extraction corpus suite per prompt change and record per-file edit counts:

1. Run the corpus suite with the current prompt:

   nix-shell -p poppler-utils --run \
     'DATABASE_URL=... SAGE_API_KEY=... cargo test --test extraction_corpus -- --ignored --nocapture'
   

2. Review the worksheet — for each fixture, the suite prints:

   • subject, question, confidence, assignee hint, spec section, drawings

3. Upload each fixture through the dashboard and click "Create RFI from extraction." Count how many fields you manually edited before submitting.

4. Record edit counts in the table below.

5. Iterate the prompt — edit extraction_system_prompt in config.toml, restart the container (ditc deploy), re-run the suite.

Calibration table

Fixture	Prompt ID	Subject edits	Question edits	Assignee edits	Total edits	Notes
rfi1_civil_conflicting
rfi2_foundation_conduits
rfi4_mat_slab_transition

Target (phase3-draft)

• 80% of vendor RFI uploads yield an extraction where the operator can click "Create RFI" with ≤2 manual edits.
• Below 50% = Phase 3 extraction is net-negative; evaluate PlanGrep pivot or drop the feature.

Prompt iteration (phase3-draft AC-5)

Edit extraction_system_prompt in config.toml and restart the container. Each prompt change gets a new prompt_id (short hash of the prompt text), stored per extraction row for attribution. The corpus suite prints the prompt_id so you can compare results across prompt versions.