Skip to content

MEMORY_AND_RAG.md

Latest commit

 

History

History
55 lines (31 loc) · 2.14 KB

MEMORY_AND_RAG.md

File metadata and controls

55 lines (31 loc) · 2.14 KB
Raw

Memory, embeddings, and RAG behavior

PostgreSQL

Connection

After authentication, db_params on the active kun.py profile is used for psycopg.connect(**db_params).

Table shape (expected)

The code issues selects and inserts against conversations with columns including id, timestamp, prompt, response.

Normalization

normalize_text applies Unicode NFKC before insert.

ChromaDB

  • Client: chromadb.Client() (ephemeral in-process in the reference build).
  • Global collection: conversations.
  • IDs: stringified PostgreSQL id.
  • Document payload: concatenation prompt: {text} response: {text}.

Hydration

If the collection is empty but SQL has rows, create_vector_db embeds historical rows using the configured embedding model.

Recall pipeline (/recall)

  1. create_queries(prompt) — Asks the configured chat model for a Python-parseable list of search strings; ast.literal_eval parses with fallback.
  2. retrieve_embeddings — Embeds each query, queries Chroma, applies classify_embedding gating.
  3. Accepted snippets become a Retrieved memory: system message; stream_response runs on the recall text.

/forget

Deletes the latest SQL row (max id) and matching Chroma id; trims two tail messages from RAM convo.

/memorize

Stores the text as a user message with assistant reply Memory specimen successfully implanted. so the pair embeds like normal dialogue.

Substring memory (recall_information)

Used on the conversational path: scans the entire table for case-insensitive substring matches in prompt or response. Cost grows with table size.

Room memory

Room maintains a separate Chroma collection (room_<slug>). It is orthogonal to the global mnemonic unless you merge exports manually.

Session exports and a longer-lived room product surface may also align with arpahls/rooms if you outgrow the bundled room.py workflow.

Backup guidance

  • Dump PostgreSQL regularly (pg_dump).
  • Treat SQL as source of truth; re-embed into Chroma if the in-process collection is lost.