Running Rereflect Fully Offline With a Local LLM

Why offline analysis matters

Customer feedback is some of the most sensitive text a company holds. It contains names, account details, candid frustrations, and sometimes data customers assumed was private. The default mode for most AI feedback tools is to ship that text to a third-party API for processing — which means your customers' words leave your control the moment they are analyzed.

For a lot of teams, that is a non-starter. Maybe you operate under a data residency requirement. Maybe your security team will not approve sending customer data to an external model provider. Maybe you simply do not want a copy of every support complaint sitting in someone else's logs.

Rereflect is self-hosted and open source, and it is built so that you never have to make that trade-off. You can run the entire analysis pipeline against a model on your own hardware. Nothing about a piece of feedback — the raw text, the inferred sentiment, the extracted pain points — ever crosses your network boundary.

BYOK, including bringing your own model

Rereflect follows a bring-your-own-key (BYOK) philosophy. There is no Rereflect-managed AI service in the middle, no per-seat AI markup, and no usage metering on our side. You connect Rereflect to whatever model you want to use, and you pay that provider directly (or, in the local case, pay nobody).

BYOK usually means pasting in an OpenAI or Anthropic key. But the same mechanism that lets you point Rereflect at a hosted provider also lets you point it at a model running on localhost. From Rereflect's perspective, a local model exposed over an OpenAI-compatible HTTP endpoint looks exactly like a remote one — it just happens to live on your own machine.

That gives you three deployment shapes, all using the same configuration surface:

• Hosted API with your own key — point Rereflect at OpenAI, Anthropic, or Google using a key you own. Best accuracy, you pay the provider per token.
• Local LLM, no key — run a model with Ollama or any OpenAI-compatible server on your infra. Keyless, private, $0 in API costs.
• No model at all — skip LLM configuration entirely and fall back to the built-in VADER analyzer, which runs locally with zero dependencies.

Pointing Rereflect at a local model

The most common offline setup uses Ollama, which runs open-weight models locally and exposes an OpenAI-compatible API. Once Ollama is running and you have pulled a model, you configure Rereflect to talk to it through the same base-URL and model-name settings you would use for any OpenAI-compatible provider.

The shape of the configuration is simple: tell Rereflect the base URL of your local endpoint and the model name to request. Because the endpoint is OpenAI-compatible, no API key is required — you can leave the key blank or set a throwaway placeholder, since the local server does not authenticate.

• Run a local server — start Ollama (or another OpenAI-compatible runtime such as llama.cpp's server, LM Studio, or vLLM) and load a model that suits your hardware.
• Set the base URL — point Rereflect's LLM base URL at your local endpoint (for Ollama, that is its local OpenAI-compatible address).
• Set the model name — specify the model you pulled, so requests ask for the right weights.
• Leave the key empty — local endpoints do not require authentication, so no API key is needed.

After that, the analysis pipeline behaves identically to a hosted setup. Feedback gets sentiment scores, pain point categories, feature request extraction, and urgency flags — all computed by the model on your own machine. The only difference is that no request ever leaves your network.

Smaller local models will not match a frontier hosted model on the hardest, most ambiguous feedback. But for the bulk of routine categorization and sentiment work, a capable local model is more than good enough — and the privacy and cost properties are unbeatable.

The free VADER fallback

Not every team wants to run a model at all. Maybe you do not have a GPU handy, or you just want to see Rereflect working on your data before you commit to any AI setup. For that, Rereflect ships with a built-in fallback: VADER.

VADER (Valence Aware Dictionary and sEntiment Reasoner) is a lexicon and rule-based sentiment analyzer. It runs entirely in-process with no model weights, no GPU, and no network calls. When no LLM is configured, Rereflect automatically uses VADER for sentiment analysis so the product still works out of the box.

It is important to be honest about what the fallback does and does not give you. VADER is a sentiment engine, not a general-purpose language model:

• What you get — fast, fully local sentiment scoring (positive / neutral / negative) on every piece of feedback, with zero configuration.
• What is reduced — the richer LLM-driven steps (nuanced pain point detection, feature request extraction, urgency reasoning) are stronger with an actual language model behind them.
• When it is enough — early-stage teams, smaller volumes, or anyone who wants an immediate, dependency-free sense of sentiment trends before wiring up a model.

The practical path most self-hosters take: start with the VADER fallback to confirm Rereflect is ingesting and scoring your feedback, then point it at a local LLM once you want the deeper categorization. Both modes keep your data on your own infrastructure.

What this costs: nothing

The offline setup has a genuinely simple cost story. There is no Rereflect subscription — the software is open source and self-hosted. There is no AI bill — a local model runs on hardware you already own, and the VADER fallback has no marginal cost at all. There is no per-seat pricing and no usage metering.

Your only real cost is the compute you choose to provision. If you run a local model on an existing server or a developer workstation, the incremental cost of analyzing feedback is effectively electricity. If you decide later that you want higher accuracy on hard cases, you can switch a single configuration value to a hosted provider with your own key — and you pay that provider directly, with no markup in between.

This is the core promise of the self-hosted, BYOK model: you own the data, you own the infrastructure, and you decide exactly how much (if anything) to spend on intelligence.

Getting started

If privacy or cost has kept you away from AI-powered feedback analysis, the offline path removes both objections. You can stand up Rereflect, point it at a local model — or just let the VADER fallback handle sentiment — and start seeing categorized, scored feedback without a single byte leaving your network.

Rereflect is open source under the MIT license. Clone it, run it on your own infrastructure, and configure the analyzer to match your privacy and budget requirements. Whether that means a local Ollama model, your own OpenAI key, or no model at all, the choice stays entirely yours.