Source: https://jtpck.com/case-studies/amc-boston-doc-chat [Case Studies](https://jtpck.com/case-studies) / AMC Boston Doc Chat # Less time digging through documents, more time on the trail. The Boston Chapter of the Appalachian Mountain Club runs on volunteers, and the knowledge they need was spread across many places. JTPCK built AMC Boston Doc Chat, a retrieval-augmented AI assistant that answers their questions in plain English and cites the document behind every answer. It runs on the web and as a native Mac app. At a glance Concept to launch ~2 wks Platforms Web + Mac Sources retrieved per answer 5 Volunteer hours saved 100s The challenge ## The answers existed. Finding them took too long. A chapter as active as AMC Boston produces a lot of documentation: policies, procedures, guides for trip leaders, committee material, and the reference material that keeps a busy schedule of outdoor activities running safely. Over the years that material ended up in many different places, written by different people in different formats. For volunteers and staff, a simple question often meant opening folder after folder, skimming long PDFs, or asking someone who might remember where the answer lived. That is time volunteers would rather spend leading trips and being outdoors, not searching for paperwork. The chapter needed one place to ask a question and get a trustworthy answer, with a way to check the source. The system ## One collection, one question box, an answer with sources. 1. 01 ### Collect Documents gathered into one library PDF, Word, text, Markdown 2. 02 ### Extract Text pulled and split into passages ~500 tokens each 3. 03 ### Embed Voyage AI embeddings in Postgres pgvector 4. 04 ### Retrieve Search by meaning for each question Top 5 passages 5. 05 ### Answer Response streamed with citations Every source clickable **Collect.** Chapter administrators upload documents to a single library and tag each one with the group that owns it. Duplicates are caught automatically, so the same policy does not show up three times under three file names. **Extract.** A background job pulls the text out of each file and splits it into passages small enough to retrieve precisely but large enough to keep their meaning. **Embed.** Each passage becomes an embedding, a numerical fingerprint of what it means, stored in the same Postgres database as the rest of the app. No separate vector database to run or pay for. **Retrieve.** A volunteer's question is embedded the same way and matched against the library by meaning, so "what do I need before leading a hike" finds the right policy even if it never uses those words. **Answer.** The language model writes the answer from the retrieved passages only and marks each sentence with a numbered citation. Clicking a citation opens the exact passage and the document it came from. The response streams in as it is written, so people are not left watching a spinner. Design choices ## Built for people who need to trust the answer. ### 1. Every answer shows its work. Volunteers make real decisions with this information, often about safety. An answer they cannot check is not good enough. Requiring a citation on every sentence, with the source one click away, keeps the assistant honest and lets people verify anything that matters. ### 2. The documents are data, not instructions. Retrieved passages are fenced off and the model is told to treat them as reference material only. A document that happens to contain instructions cannot redirect the assistant. This kind of prompt-injection hygiene is easy to skip and matters more as a library grows. ### 3. Private by default. Accounts are required, conversations belong to the person who started them, and a source passage is only viewable from a conversation that actually cited it. Volunteers can choose to share a useful conversation publicly, but nothing is shared unless they do. ### 4. Meet people where they already work. The assistant runs in any browser and as a native Mac app on the App Store, so it can sit in the dock next to email and the calendar rather than in a forgotten bookmark. ### 5. Simple to run for a volunteer organization. The whole system is one Rails application on one Postgres database, with background jobs and real-time updates handled in the database too. No Redis, no separate vector store, no fleet of services. A nonprofit should not need a platform team to keep its document assistant running. The outcome ## Questions answered in seconds, with the source attached. Chapter volunteers now ask their question in one place instead of hunting through scattered files or waiting on someone who knows where things are. The answer arrives in seconds with the source attached, and when a policy changes, updating the library updates every answer. The result is less volunteer time spent wading through documentation and more time doing what they signed up for: getting people outdoors. Try it at [chat.amcboston.org](https://chat.amcboston.org) or [on the Mac App Store](https://apps.apple.com/us/app/amc-boston-doc-chat/id6759003243). Beyond the trail ## Every organization has this problem. Swap trip-leader guides for an employee handbook, a policy manual, or years of project files and the problem is the same: the answer exists, but it is buried, and the people who need it waste time looking or give up and guess. - **Internal knowledge assistants.** Policies, procedures, and HR questions answered with a citation. [RAG AI systems](https://jtpck.com/services/rag-ai-systems) - **Nonprofits and member organizations.** Give volunteers and staff the institutional knowledge that usually lives in a few people's heads. - **Professional firms.** Practice manuals, templates, and past work, searchable by meaning. [AI for law firms](https://jtpck.com/ai-consulting/law-firms) - **Sensitive material.** The same architecture can run entirely on your own hardware. [On-premise AI](https://jtpck.com/services/on-premise-ai) Get started ## Are your documents scattered too? JTPCK builds custom RAG systems for businesses and organizations in Boston, across New England, and remotely. Tell us where your knowledge lives and who needs it. [Discuss a project with JTPCK](https://jtpck.com/contact) [RAG AI systems](https://jtpck.com/services/rag-ai-systems)