Add bounded multi-turn chat memory (prompt.py)

prompt.py

@@ -13,7 +13,16 @@ SYSTEM_PROMPT = (
 )
 
 # Keep the context budget modest so generation stays fast on ZeroGPU.
-MAX_CONTEXT_CHARS = 6000
+MAX_CONTEXT_CHARS = 5000        # trimmed a bit since history now shares the prompt
+
+# Multi-turn memory bounds (keep prompts — and ZeroGPU time — bounded).
+MAX_HISTORY_TURNS = 4          # default # of prior user+assistant exchanges kept
+MAX_HISTORY_CHARS = 6000       # hard cap on total history text fed to the model
+
+# Marker that app.py appends after the model's answer; everything from here on is
+# our validation/sources/notes decoration and must be stripped before the answer
+# is fed back as conversation history.
+VALIDATION_DELIM = "\n\n---\n**Validation:**"
 
 
 def _format_context(hits: list[Hit]) -> str:
@@ -29,13 +38,56 @@ def _format_context(hits: list[Hit]) -> str:
     return "\n\n".join(blocks)
 
 
+def _clean_assistant(content: str) -> str:
+    """Strip our validation/sources/notes so only the model's answer remains."""
+    return content.split(VALIDATION_DELIM, 1)[0].rstrip()
+
+
+def _normalize(history) -> list[dict]:
+    """Flatten gradio history (messages OR tuples format) to role/content dicts."""
+    out: list[dict] = []
+    for item in history or []:
+        if isinstance(item, dict) and "role" in item:
+            role, content = item.get("role"), item.get("content", "")
+            if not isinstance(content, str):     # skip file/component messages
+                continue
+            if role == "assistant":
+                content = _clean_assistant(content)
+            if content.strip():
+                out.append({"role": role, "content": content})
+        elif isinstance(item, (list, tuple)) and len(item) == 2:
+            u, a = item
+            if isinstance(u, str) and u.strip():
+                out.append({"role": "user", "content": u})
+            if isinstance(a, str) and a.strip():
+                out.append({"role": "assistant", "content": _clean_assistant(a)})
+    return out
+
+
+def _prepare_history(history, max_turns: int) -> list[dict]:
+    """Last `max_turns` exchanges, sanitized and char-bounded, starting on a user turn."""
+    if max_turns <= 0:                       # 0 = single-turn ([-0:] would keep ALL)
+        return []
+    msgs = _normalize(history)[-2 * max_turns:]
+    # Apply the char budget from the most recent message backwards.
+    total, bounded = 0, []
+    for m in reversed(msgs):
+        total += len(m["content"])
+        if bounded and total > MAX_HISTORY_CHARS:
+            break
+        bounded.append(m)
+    bounded.reverse()
+    while bounded and bounded[0]["role"] != "user":   # don't start on an assistant turn
+        bounded.pop(0)
+    return bounded
+
+
 def build_messages(question: str, hits: list[Hit],
-                   history: list[dict] | None = None) -> list[dict]:
-    """Build chat-template messages for the generator."""
+                   history=None, max_turns: int = MAX_HISTORY_TURNS) -> list[dict]:
+    """Build chat-template messages: system + bounded history + context+question."""
     context = _format_context(hits)
     messages: list[dict] = [{"role": "system", "content": SYSTEM_PROMPT}]
-    if history:
-        messages.extend(history)
+    messages.extend(_prepare_history(history, max_turns))
     user = question if not context else (
         f"Reference GDScript snippets from a curated Godot corpus:\n\n"
         f"{context}\n\n---\n\nQuestion: {question}"