spec(M4): RAG Knowledge Base — contextual retrieval and multi-turn chat memory #8

New issue

Open

opened 2026-05-19 07:46:44 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-05-19 07:46:44 -07:00

Owner

Goal

Robin's M2 chat answers from the LLM's training data only. M4 adds two things: a curated migration knowledge base Robin can retrieve from, and multi-turn conversation memory so Robin remembers what you said earlier in the session.

Knowledge base content

The knowledge base is a set of structured documents covering:

Software equivalents: macOS → Arch, Windows → Debian, Debian → Arch (see issue #4)
Common error patterns and their plain-language explanations
Distro-specific gotchas (AUR mechanics, pacman lock, apt held packages, etc.)
Terminal command translations (brew → pacman/paru, apt → pacman, etc.)
File path and config location differences between distros
Hardware/driver questions common at migration time (wifi, GPU, Bluetooth)

Content is seeded from the real friction points Xander and Alan's partner encounter during their migrations — this is the primary acceptance test.

Technical approach

Same chunk/embed/retrieve pipeline as cf-core (shared with Turnstone). Documents are chunked, embedded locally, stored in a SQLite vector table. At chat time, the top-k chunks most relevant to the user's question are prepended to the system prompt.

All embedding runs locally — no data leaves the machine.

Multi-turn memory

M2 sends only the current message. M4 keeps a rolling window of the conversation (last N turns) so Robin can answer follow-up questions without losing context: "How do I install that?" after "What's the Arch equivalent of Homebrew?" should work correctly.

Relationship to issue #4

Linux-to-Linux migration paths (Debian/Fedora → Arch) land here. The knowledge base is where distro-specific translation content lives. See issue #4 for the full spec.

Open questions

Xander: What are the questions you've actually had to look up since switching to CachyOS? Those are the first entries in the knowledge base.
How many turns of memory is useful before the context gets noisy? (Starting assumption: last 10 turns)
Should Robin indicate when it is using the knowledge base vs. answering from general LLM knowledge? (Transparency principle)
Should the knowledge base be community-editable (MIT layer) while the retrieval and chat logic stays BSL?

Dependencies

M1, M2, M3 merged
cf-core RAG pipeline (same as Turnstone#21)
issue #4 (Linux-to-Linux paths)

## Goal Robin's M2 chat answers from the LLM's training data only. M4 adds two things: a curated migration knowledge base Robin can retrieve from, and multi-turn conversation memory so Robin remembers what you said earlier in the session. ## Knowledge base content The knowledge base is a set of structured documents covering: - Software equivalents: macOS → Arch, Windows → Debian, Debian → Arch (see issue #4) - Common error patterns and their plain-language explanations - Distro-specific gotchas (AUR mechanics, pacman lock, apt held packages, etc.) - Terminal command translations (brew → pacman/paru, apt → pacman, etc.) - File path and config location differences between distros - Hardware/driver questions common at migration time (wifi, GPU, Bluetooth) Content is seeded from the real friction points Xander and Alan's partner encounter during their migrations — this is the primary acceptance test. ## Technical approach Same chunk/embed/retrieve pipeline as cf-core (shared with Turnstone). Documents are chunked, embedded locally, stored in a SQLite vector table. At chat time, the top-k chunks most relevant to the user's question are prepended to the system prompt. All embedding runs locally — no data leaves the machine. ## Multi-turn memory M2 sends only the current message. M4 keeps a rolling window of the conversation (last N turns) so Robin can answer follow-up questions without losing context: "How do I install that?" after "What's the Arch equivalent of Homebrew?" should work correctly. ## Relationship to issue #4 Linux-to-Linux migration paths (Debian/Fedora → Arch) land here. The knowledge base is where distro-specific translation content lives. See issue #4 for the full spec. ## Open questions - Xander: What are the questions you've actually had to look up since switching to CachyOS? Those are the first entries in the knowledge base. - How many turns of memory is useful before the context gets noisy? (Starting assumption: last 10 turns) - Should Robin indicate when it is using the knowledge base vs. answering from general LLM knowledge? (Transparency principle) - Should the knowledge base be community-editable (MIT layer) while the retrieval and chat logic stays BSL? ## Dependencies - M1, M2, M3 merged - cf-core RAG pipeline (same as Turnstone#21) - issue #4 (Linux-to-Linux paths)

pyr0ball added this to the M4: RAG Knowledge Base milestone 2026-05-19 07:46:44 -07:00

No labels

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference: Circuit-Forge/robin#8

No description provided.