Pathway

Support that knows its sources.

Pathway is a RAG chatbot for non-clinical pharmaceutical access and patient-support work. Choose a fictional case, ask a question, prepare an email or case briefing, and inspect the documents behind the answer.

Try the live demo · Explore the knowledge · View evaluation

Why I built Pathway

I kept seeing avatar-based AI agents marketed with expensive licenses, and I wanted to try building one myself. For this experiment, I started with a chatbot that uses retrieval-augmented generation (RAG): it finds relevant passages in a selected knowledge base, gives them to the language model, and shows the sources alongside the answer.

I try to learn a different tool with each project. Here, I chose Render to host the frontend and application server, and Supabase for PostgreSQL and pgvector. Building and deploying Pathway gave me a way to learn how those pieces work together.

The three support roles have illustrative stock portraits. The interaction happens through Chat and Email, with source inspection and drafts for review. The people in the portraits are not staff or endorsers; photo credits and licensing are included.

The workspace

Case context stays on the left; conversation and drafts have their own space on the right. On mobile, Case context opens a keyboard-accessible drawer. The composer stays visible beneath the conversation.

• Understand the case. See your role, the recorded case snapshot, where things are stuck, and the selected Knowledge Pack’s version, topics and applicable document count.
• Ask naturally. Follow up with questions such as “Why?” or “What remains unknown?” Pathway uses recent conversation within the same knowledge boundary to resolve references.
• Work in Chat or Email. Create office follow-ups, case briefings and patient-friendly administrative messages. Copy a draft or ask Pathway to refine it. Drafts are generated for review and never sent.
• Inspect the evidence. Sources shows exact cited passages with document titles, sections and versions. Grounding explains the contextual query, retrieval ranks and passages used to prepare the answer.
• See work in progress. Live activity shows knowledge search, retrieved document counts, answer preparation and source-reference checks. Elapsed time is visible while waiting; completed activity remains available above the answer.
• Keep control. Copy, regenerate, give feedback or retry a failed request. Conversations, source snapshots and completed activity survive reload within the browser session.

The sidebar describes the scenario’s recorded snapshot. Conversation text does not change case status. Switching knowledge updates the current context while earlier answers retain their original case and sources.

See sources, activity and mobile views

Exact supporting passages

Visible activity

Mobile conversation and case context

Conversation screenshots show the current interface replaying retained synthetic answers locally, without new model calls. The scenario cards below were captured from the public demo. Capture details.

Three scenarios

Scenario	Your role	Recorded blocker	Useful output
Documentation support · ALD-1042	Avery Chen, HCP office access coordinator	The submitted package has an unsigned office note; the signed note and signature page are missing.	An office email requesting the signed note.
Coverage & access · ACC-2086	Morgan Lee, field reimbursement manager	An unreadable Juniper Plan member identifier blocks benefit verification.	An HCP-office briefing or follow-up email.
Specialty pharmacy fulfillment · FUL-3091	Jordan Rivera, patient-support case manager	The consent signature is missing; intake is recorded, but no dispense release or carrier handoff is recorded.	A patient-friendly administrative status message.

All case identities, organizations, plans, cases and documents are fictional. Each Knowledge Pack contains 14 documents: 12 applicable sources, one superseded source and one record for another case. The complete corpus contains 42 documents and 110 passages, including practical checklists, troubleshooting, handoff worksheets and communication examples. The guides explain fictional procedures without adding events to the recorded cases. Excluded documents are inspectable in the knowledge browser but cannot support a new answer. See the Knowledge Pack guide and verification.

Try a conversation

Use Walkthrough inside a conversation or on Knowledge for a short guide to the controls. A light fade and teal outline keep the page visible. Tours are optional, work with a keyboard, and never submit a question or change the selected knowledge.

1. Open the demo and choose HCP office documentation support.
2. Ask “What document is missing?”, then “Why is it being requested?”
3. Ask “Draft a short email requesting the document.” Follow with “Make it warmer and shorter.”
4. Open Sources to read the supporting passages, or Grounding to inspect retrieval.
5. Switch Knowledge Packs to explore another case and its evidence boundary.

How it works

flowchart LR
    A[Question and recent same-pack context] --> B[Current pack and case filters]
    B --> C[Hybrid retrieval]
    D[(PostgreSQL and pgvector)] --> C
    C --> E[OpenAI GPT-5 mini]
    E --> F[Response-format and source-ID validation]
    F --> G[Saved answer, draft, sources and activity]
    G --> H[Chat and Email workspace]

Loading

Layer	Implementation
Interface	Semantic HTML, CSS and browser JavaScript; responsive context drawer and independent conversation scrolling.
Server	Node.js 24 and TypeScript, with session-scoped HTTP APIs.
Generation	OpenAI GPT-5 mini through the Responses API, with low reasoning effort and structured outputs.
Retrieval	PostgreSQL/pgvector hybrid vector and lexical ranking; OpenAI text-embedding-3-small embeddings cached persistently.
Evidence	Current-source, pack and case filtering; validated citation membership and saved exact source snapshots.
Progress	Persisted application events streamed as NDJSON, with bounded status polling when the hosting proxy buffers updates. No additional model calls.
Hosting	Render static frontend and Node backend, with a dedicated Supabase PostgreSQL database.

Generation uses retrieved evidence and explicit scenario context. Conversation history helps resolve references; it does not establish new facts. A malformed response can receive one bounded format repair using the same evidence. API keys stay on the server, and provider request budgets are enforced persistently.

Run locally

Requires Node.js 24.12+, PostgreSQL 18.6 and pgvector 0.8.6. The local launcher supports the documented macOS/Homebrew installation; see the database setup guide for pinned binaries and connection configuration.

npm ci
npm run db:start
npm run db:migrate
npm run rag:seed
npm run build
cp .env.example .env

In your ignored .env file, set OPENAI_API_KEY to your own key and PATHWAY_LIVE_GENERATION=enabled to enable answers. Keep PATHWAY_RAG_MODEL=gpt-5-mini, PATHWAY_RAG_REASONING=low, and positive provider request limits. Live answers and uncached embeddings use your OpenAI account’s API budget.

node --env-file=.env dist/src/app/server.js

Open localhost:3000. Without live generation configured, the catalog and interface remain available, but Pathway cannot produce a live answer. The first enabled request indexes any missing corpus embeddings.

Verification

npm run typecheck
npm test
npm run test:rag
npm run build
node scripts/verify-hygiene.ts

The RAG suite requires the migrated local database. Automated checks use synthetic provider responses and make no paid model calls. The documented activity release passed 118 automated tests and 138 browser checks, including desktop/mobile layouts, citations, copy, refinement, retries, knowledge switching, progress and reload.

The retained 36-turn synthetic conversation evaluation of Knowledge Pack version 1.0 completed all turns, with 98.61% expected-passage recall and 115/115 cited IDs present in retrieval. Its semantic review found 34/36 complete criterion matches; the two partial results and targeted follow-up checks are documented. Citation membership alone does not prove every phrase is supported, and these results are not independent clinical validation.

• Conversation evaluation and limitations
• Workspace behavior and browser checks
• Activity, persistence and public verification
• Canonical scenario and knowledge definitions
• Response and conversation contracts

Demonstration boundaries

Use the supplied synthetic cases. Pathway does not provide clinical advice, make payer or eligibility decisions, promise dispensing or delivery, or perform real outreach. Review generated drafts before using their wording. Sessions expire after four hours; this portfolio demonstration is not a production healthcare system.