Pūnaha 1.0.0: Garden Club knowledge chat, full local walkthrough Captured 19 November 2026 on an installed, registered Windows Pūnaha 1.0.0 build. Nika narration added 20 November 2026. The recording opens in a separate Garden Club Tutorial tenant. Existing Gemma E4B chat and EmbeddingGemma GGUF files are connected and enabled. Their download and GGUF upload are not shown because the files were already installed. The model files remain shared; the new tenant gets its own connection and deployment records. A new Garden club knowledge database is created with Semantic Chunking, EmbeddingGemma, 768 dimensions and no processing model. The fictional patchwork-garden-club.txt practice file is genuinely selected in the app and uploaded. Document processing reaches Embedding completed and 100%. A retrieval search for the club meeting finds the source passage about Sunday at 9:30 am in the Glasshouse Room. The Answer from approved knowledge workflow template is created and configured with the new database, Balanced retrieval, a prompt that requires support for every requested fact, and Gemma E4B at temperature zero. A Condition checks the model output for the exact missing-answer sentence. Its TRUE route returns only that sentence; its FALSE route returns the supported answer. Draft tests answer the meeting question and the seven-day hand-tool borrowing question. They return exactly "I can't find an answer to your question." for the undocumented annual fee and for a question combining a known meeting time with that missing fee. After local workflow publication, Chat answers the meeting question, refuses the missing annual fee, answers the hand-tool borrowing question, and answers the meeting question again. The model labels its supported answers "Source: Document"; this label does not substitute for checking the retrieved passage. These observed responses do not guarantee future model behaviour or qualify other installations. The recording consists of two real browser captures joined where the recording browser disconnected. The sign-in capture was discarded. The first 25 seconds of the first retained capture were removed, and the workstation label was concealed in the processing screen. Waiting periods remain in the video. No synthetic application screens or fabricated answers were added. Spoken narration: AI-generated Nika voice in Google AI Studio, natural style. 00:00:00.000 We begin in a separate Garden Club Tutorial tenant. The Gemma chat and EmbeddingGemma files are already on this server. This video connects their existing deployments to the tenant; it does not repeat the model download or GGUF upload. 00:00:30.000 Check that both local model entries are enabled. Gemma E4B writes the chat answers. EmbeddingGemma turns text into vectors for search. They have different jobs; an embedding model cannot answer the user in Chat. 00:01:00.000 Now create the Garden club knowledge database. Choose Semantic Chunking and EmbeddingGemma, with 768 dimensions. There is no separate processing model in this example. These settings prepare the practice document for retrieval. 00:01:45.000 Select and upload the fictional Patchwork Garden Club text file. Wait until document processing and embedding reach completion. Then search for the meeting and inspect the retrieved passage: Sunday at nine thirty in the Glasshouse Room. The match score measures similarity, not answer confidence. 00:02:30.000 Next, create a workflow from the Answer from approved knowledge template. Give it the Garden club knowledge chat name. This is a draft until we test and publish it inside Pūnaha. 00:03:15.000 On the canvas, connect the knowledge retrieval stage to the new database. The prompt must use the retrieved passage and require support for every fact in the answer. Keep the required Protection gateway in place. 00:03:50.000 Choose the Gemma E4B chat deployment for answer generation. Set temperature to zero and response detail to Balanced. Save the draft, check readiness, then test with the garden club meeting question. 00:04:30.000 The draft run is using the local model, so give it time to complete. When the meeting answer appears, compare it with the retrieved source passage. A source label by itself is not proof that the wording is supported. 00:05:05.000 Now add a Condition after the model. It checks whether the generated text contains the exact missing-answer sentence: I can't find an answer to your question. This gives the workflow two routes. 00:05:45.000 Configure the TRUE route to return only that refusal sentence. Configure the FALSE route to return the supported model answer. This keeps an unknown fact from being dressed up as a confident answer. 00:06:30.000 Review the complete path on the canvas. Retrieval feeds the grounded prompt, Gemma generates the response, and the Condition sends it to the appropriate final output. Save each setting before testing again. 00:07:20.000 Here the supported-answer output is connected to the model's response. Check the selected connection and the module references carefully. The final route should return the answer from this run, not a fixed example sentence. 00:08:05.000 Test the meeting question first. The practice document says the club meets every Sunday at nine thirty in the Glasshouse Room. Inspect the completed run and compare that answer with the actual retrieved text. 00:08:45.000 Now ask for the annual membership fee. The document gives no fee, so the expected response is the exact refusal: I can't find an answer to your question. A failed run is different from a correct refusal, so check the run status as well. 00:09:30.000 Try a mixed question that asks for both the known meeting time and the missing fee. Because the source cannot support the complete answer, the workflow should refuse rather than invent a fee. 00:10:15.000 Test another supported detail: how long members may borrow hand tools. The practice file says seven days. Check the answer and the source passage, then confirm the missing-information tests still pass. 00:11:05.000 With the draft checks complete, publish version one inside Pūnaha. Chat can now use that tested workflow. 00:11:25.000 Open Chat in the same tenant and select Garden club knowledge chat. Ask when and where the club meets. The published workflow answers from the practice document; verify Sunday, nine thirty, and the Glasshouse Room. 00:12:10.000 Ask about the annual membership fee next. The published Chat should use the same exact refusal because no fee is in the source. Do not treat a source label or a similarity percentage as independent proof. 00:12:55.000 Finally, ask about hand-tool borrowing and repeat the meeting question. The recorded replies are supported by the practice file, but model responses can vary. Check every answer against the retrieved passage on your own installation.