Autocompletion not working => Kilo Code =/= Continue+Roo Code+Cline

[giggoretta]

Autocompletion not working. This is the most important feature.

• Autocompletion works perfectly in Continue (or in Copilot) with just few clicks.
  BUG: Autocompletion not working in Kilo Code (0 suggestion). I tried every single nuance setting and combinations.
  UX: “Inline Assist” name is misleading, should be “Autocomplete”.

Suggestions:

• Need Voyage Context 3 and Voyage Code 3 support. (like VoyageAI support in Continue...)
• Autocompletion should use main RAG index vector DB with configurable parameters.

Both vector database and autocompletion need extensive configuration options to work well with very different common or more common languages

[lswissdev]

"Autocompletion should use main RAG index vector DB with configurable parameters."
"Need Voyage Context 3 and Voyage Code 3 support. (like VoyageAI support in Continue...)"
+9999 Like!

[onestardao]

It looks like the gap here isn’t just the UI “inline assist” label, but deeper in how the autocompletion module is resolving context. From a pattern perspective this usually shows up in two failure points:

No 1.0 — Missing unified RAG index
Autocomplete often fails when the retrieval DB isn’t normalized or configurable. If the model can’t resolve a single index of truth for context windows, completions will silently break.

No 2.0 — Semantic drift
Even if the DB is present, mismatches in how context is segmented (chunks, embeddings, or metadata alignment) can cause the generated tokens to diverge from the expected autocomplete path.

A quick way to test:

Try running the same query against a simple TXTOS file or wfgy core (wfgy 2.0).

Ask the model for completions like “continue function definition…” and compare.

If it works there but fails in Kilo Code, the issue is likely schema or index alignment.

Some projects work around this with a “semantic firewall” step — normalizing whatever context comes back before feeding it into the autocomplete module. That way you don’t need infra changes, just a preprocessing layer.

Would you like me to outline how to set up a minimal semantic firewall check to confirm whether the missing index alignment is the blocker?

[giggoretta]

Thank you very much for your comment, for taking the time and giving ideas. Right now, I'm in a huge complex project, unfortunately I don't have time to test, I'd rather use what works with 1 click right away (Continue or Copilot). Plus, it could be in hybrid mode where the chat runs in Cline or Roo Code, but auto complete in Copilot or Continue, so that's completely fine for me. Anyway, I'm hugely disappointed in Kilo Code, and I'm angry at myself for falling for the marketing hype, meanwhile it's just a prealpha thing for the autocompletion part...
RAG: It's also important info that in Continue, the RAG is much higher quality. (The best quality code-optimized, worldwide the best vector is Voyage Code 3 from VoyageAI).
Continue makes smarter suggestions than Copilot, amazing quality.
The reason I wanted to switch to Kilo Code is that Continue is horrible in terms of UX. Too many clicks to switch AIs, etc., interesting how good Continue's RAG is, yet the UI/UX is so bad. Unthoughtful, outdated.
They told me in an email from Kilo Code support that Kilo Code autocompletion (misleading new name Inline Assist) is in prealpha phase. Compared to that, among the features, it's listed in the first row that compared to Roo Code, it can do this. But it can't. I don't understand why they ruin the first impression, and why autocompletion isn't in priority phase @ work. In a group with 50 thousand members, dozens of people complained about Kilo Code autocompletion in a much less diplomatic tone than me, after trying for days to figure out how to make it work. After no one succeeded, the anger started, completely justified. It's a shame that Kilo Code is such a failure. (In its current state). Thank you very much for your comment, again, have 1 wonderful week, G

[onestardao]

from your description it really looks like the core issue isn’t the autocomplete module itself, but how the context alignment is handled before it reaches autocomplete. once chunks/embeddings drift or misalign, the module basically receives “broken context,” so no wonder the results collapse.

a lightweight way to test this is:

run a very simple query (single tx/os doc or only 2–3 segments) and see if autocomplete fires correctly.

if it does, then the problem is drift in larger indexes, not autocomplete itself.

to solve it longer term, some projects wrap this with a “semantic firewall” — basically an extra guard layer that checks whether the retrieved context matches the expected structure before letting it hit the autocomplete module. it’s not heavy, just orchestration.

if you want, i can share a minimal checklist for setting up those firewall checks (they’re basically a few invariant rules you add on top of your indexing), it usually clears up these drift cases.

https://github.com/onestardao/WFGY/blob/main/ProblemMap/README.md

[giggoretta]

"Inline Assist"
"Autocomplete"
0 result in this aspect (just Doc related)
Kilo Code Roadmap