Little Kiln Support Copilot
A local AI customer-support assistant for independent creators, built on a RAG pipeline. It uses open-source models and vector retrieval to generate draft replies, with no API costs.
Timeline (last 90 days)
Little Kiln Support Copilot is a lightweight retrieval-augmented generation (RAG) system based on local open-source models, written in Python, running two open-weight models, EmbeddingGemma and Gemma 4 E2B, locally via Ollama.
In the Little Kiln Support Copilot tests, relevant historical cases scored above 0.63 in cosine similarity and noise cases below 0.58, so the author raised the retrieval threshold from 0.5 to 0.6.
After lowering temperature to 0.2 and instructing in the prompt to apply the solution from the most similar case, the model's draft for "the vase arrived broken and I want a refund" stably adopted the damaged-item policy (free replacement within 2 days, no need to return the damaged item, provide a photo of the damage) instead of incorrectly applying the no-reason returns policy.
When the retrieval threshold was set to 0.5, a new question about a teapot pulled in irrelevant cases, causing Gemma to claim the shop did not have teapots (fabricated information); after raising the threshold to 0.6, the model states that the shop owner needs to confirm when there are no relevant cases.
When using paid third-party AI APIs, customer service costs grow linearly with the number of customer messages; whereas running the model locally, replying to 20 emails a day incurs no API fees.
Little Kiln Support Copilot is currently a command-line demo that demonstrates its capabilities through three real cases: a broken vase triggering the replacement policy, a delayed package citing the SORRY10 discount code, and an uncovered question being honestly handed over to the shop owner.
In the delayed package case, the model's draft leaked internal phrasing such as "in the past we have offered...", proving that human review of every draft is necessary.
The code of Little Kiln Support Copilot is open-sourced on GitHub and has been submitted to the second week's "side project" track of the Hacktoberfest 2026 open-source AI challenge, competing for the "Best Use of Gemma" and "Best Use of Tiger Data" awards.
For the customer service draft scenario of a one-person shop, a larger closed-source model might write more fluent text, but a free, private, local model is the more cost-effective choice, because the drafts require human review.
All Facts (9)
For the customer service draft scenario of a one-person shop, a larger closed-source model might write more fluent text, but a free, private, local model is the more cost-effective choice, because the drafts require human review.
50%UnverifiedIn the Little Kiln Support Copilot tests, relevant historical cases scored above 0.63 in cosine similarity and noise cases below 0.58, so the author raised the retrieval threshold from 0.5 to 0.6.
50%UnverifiedAfter lowering temperature to 0.2 and instructing in the prompt to apply the solution from the most similar case, the model's draft for "the vase arrived broken and I want a refund" stably adopted the damaged-item policy (free replacement within 2 days, no need to return the damaged item, provide a photo of the damage) instead of incorrectly applying the no-reason returns policy.
50%UnverifiedWhen the retrieval threshold was set to 0.5, a new question about a teapot pulled in irrelevant cases, causing Gemma to claim the shop did not have teapots (fabricated information); after raising the threshold to 0.6, the model states that the shop owner needs to confirm when there are no relevant cases.
50%UnverifiedWhen using paid third-party AI APIs, customer service costs grow linearly with the number of customer messages; whereas running the model locally, replying to 20 emails a day incurs no API fees.
50%UnverifiedLittle Kiln Support Copilot is currently a command-line demo that demonstrates its capabilities through three real cases: a broken vase triggering the replacement policy, a delayed package citing the SORRY10 discount code, and an uncovered question being honestly handed over to the shop owner.
50%UnverifiedIn the delayed package case, the model's draft leaked internal phrasing such as "in the past we have offered...", proving that human review of every draft is necessary.
50%UnverifiedLittle Kiln Support Copilot is a lightweight retrieval-augmented generation (RAG) system based on local open-source models, written in Python, running two open-weight models, EmbeddingGemma and Gemma 4 E2B, locally via Ollama.
50%UnverifiedThe code of Little Kiln Support Copilot is open-sourced on GitHub and has been submitted to the second week's "side project" track of the Hacktoberfest 2026 open-source AI challenge, competing for the "Best Use of Gemma" and "Best Use of Tiger Data" awards.
50%