A recurring builder problem is not finding an IP geolocation database; it is deciding which option is defensible for a specific project. A V2EX discussion dated 2026-11-13 asks which open-source or commercial database is useful for province and city lookup, and whether a paid option offers good value. The replies mention `geoip-cn` on GitHub, a third-party API, and ipip.net, but they do not establish accuracy, coverage, current pricing, or suitability.
That gap is a useful product opportunity. Instead of pretending to provide a live lookup service or an accuracy benchmark, build a hosted comparison worksheet that records evidence, applies transparent scores, and makes uncertainty visible.
What decision does the market signal?
The immediate need is a repeatable way to compare IP-to-city data sources when the available discussion does not settle the choice. The market evidence contains a direct open-source-versus-commercial question, two short replies, and 470 views. That is enough to identify unresolved demand, but not enough to recommend a vendor.
The same evidence also shows why trust matters. Another V2EX post describes AI-generated networking claims that included invented companies, domains, prices, and configuration paths. The author explicitly labels statements such as "90%" regional support, monthly prices, and menu locations as unverified or contradicted. For a database comparison tool, that is a warning: an attractive score is not evidence unless the underlying observation is recorded.
The worksheet should therefore answer a narrower question: "Which candidate best fits the evidence and constraints we have recorded? It should not answer "Which provider is objectively most accurate?
Accuracy benchmarks, vendor pricing, live lookup results, and customer outcomes remain unverified from the supplied evidence.
Why should the worksheet separate evidence?
The worksheet should keep observations, assumptions, and calculated scores in separate fields. This prevents a plausible AI-written note from silently becoming a product decision.
A practical record can include:
- Candidate name and type: open-source, commercial, or unknown.
- Geographic scope claimed by the operator, recorded as an observation rather than a verified fact.
- Data access method, such as a local dataset or third-party API, only when the operator has documented it.
- License or usage constraint, marked unverified until checked against the relevant documentation.
- Price, refresh date, and coverage statements, each with an evidence status.
- Evidence note, source date, reviewer, and confidence level.
- Decision score and the individual factors that produced it.
SQLite is a good fit for this first version because the project needs structured records and editable review history, not a claimed global data service. Store candidates, evidence items, scoring criteria, and review decisions in separate tables. Keep the original note beside the normalized value so a reviewer can see whether a score came from a documented fact, a team assumption, or an unresolved question.
A useful rule is to make every unverified field visible in the interface. Do not replace an unknown price with zero, and do not treat a missing accuracy test as a failed test. Those are different states.
How can you build the first version?
Build a small HTML interface, a Python scoring function, and a SQLite-backed record store. The interface can contain a candidate table, an evidence editor, a scoring panel, and an exportable decision summary.
Use AI coding to draft the data model, form validation, and test cases, but keep the evidence boundary in the prompts and in the review workflow. Ask the assistant to label unsupported statements as `unverified`, preserve the original wording, and explain every score calculation.
The scoring model should be deliberately boring. For example, let a reviewer assign a value from 0 to 3 for criteria such as documented access, project fit, operational burden, and evidence completeness. Python can calculate a weighted total and return both the total and a breakdown. The exact weights are a project decision; the supplied market evidence does not justify a universal weighting scheme.
A minimal calculation might look like this:
def weighted_score(values, weights):
total = 0.0
details = {}
for key, weight in weights.items():
value = values.get(key, 0)
details[key] = value * weight
total += details[key]
return {"total": total, "details": details}You can run Python online while iterating on the scoring rules and sample records. Keep the function deterministic: the same evidence and weights should produce the same result. That makes a decision reviewable and makes later changes easier to audit.
The front end can be a plain HTML application with local form interactions and a clear status display. You can run HTML online to test the worksheet before publishing it. VicroCode's confirmed scope supports online execution, hosting, and publishing for HTML web apps, plus online Python execution, SQLite databases, file management, and project sharing. This design does not require claiming support for another language, framework, cloud service, or deployment method.
Where should AI assist the comparison?
AI should assist with organization and questions, not manufacture evidence. A useful comparison assistant can turn a reviewer note into structured fields, detect missing values, suggest follow-up questions, and draft a concise explanation of why two candidates received different scores.
For example, the assistant can flag these problems:
- A candidate is described as "accurate" without a test record.
- A commercial option is called cost-effective without a price or workload assumption.
- A city-level capability is inferred from a general IP database description.
- A source is named but its date and verification status are missing.
- A score changes even though the evidence record did not.
Model Center APIs can support notes for models already available through the platform, while an in-platform tool call can request a Python recalculation or fetch a saved SQLite record. Keep the tool boundary narrow. The model may propose a clarification or format text, but the score should come from the stored criteria and the Python function.
This separation matters because the evidence includes a post warning that AI output had invented domains, prices, and technical settings. An AI comparison note that sounds confident is not a substitute for a checked record.
What should remain explicitly unverified?
Live IP lookup, accuracy rankings, vendor pricing, and customer outcomes should remain unverified in this version. Nothing in the supplied evidence provides a controlled comparison, current pricing table, benchmark dataset, or confirmed production result.
The tool can still be useful without those claims. It can record a team's own test cases when they later become available, store the date and method of each test, and compare candidates against the team's stated requirements. It can also show "not tested" separately from "failed" and "unknown price" separately from "free."
Do not add a live lookup feature unless the required data source, API terms, and platform fit are separately confirmed. The confirmed VicroCode capabilities cover API Endpoint Hosting and in-platform tool calls, but the supplied evidence does not confirm any particular IP geolocation provider, lookup endpoint, data license, or accuracy guarantee. The safe product boundary is therefore a comparison worksheet, not a claimed geolocation service.
How can a small team operate it?
A small team can use the tool as a decision log rather than a one-time ranking page. Start with one project, enter the candidates mentioned in the discussion only as leads, and mark every unverified property before assigning a score.
Reviewers can then add evidence in stages:
- Record the requirement, such as province-level or city-level output.
- Add each candidate and preserve its source wording.
- Mark unknown licensing, price, coverage, and accuracy fields.
- Run the same scoring function with documented weights.
- Save the decision, reviewer, date, and unresolved questions.
- Revisit the score when new evidence arrives.
Publishing the worksheet can make the method shareable, but publishing it does not turn a community mention into validation. The first V2EX discussion had 470 views and only two replies in the supplied extract. That makes it a useful signal of an unresolved choice, not a representative market survey.
A later version could add project sharing, monetization, or an API endpoint for internal scoring, all within the confirmed platform capabilities. It should still avoid implying that the tool knows which database is best until the team has collected comparable evidence.
FAQ
Is this tool a live IP lookup service?
No. The proposed version is a hosted comparison worksheet with Python scoring, SQLite records, and review notes. Live lookup and provider accuracy are unverified from the supplied evidence.
Can AI choose the best database automatically?
No. AI can structure notes, identify missing evidence, and explain stored scores, but it should not invent prices, coverage, accuracy, or vendor capabilities. The final score should come from explicit criteria and recorded inputs.
What should an operator test first?
Test the evidence workflow first: candidate records, unverified labels, reproducible scoring, and saved decisions. Add live data testing only after a suitable provider, access method, licensing terms, and benchmark method have been confirmed.