almost finished
This commit is contained in:
@@ -37,48 +37,7 @@ with open(project_root / "data" / "processed" / "parent_lookup.json") as f:
|
||||
parent_lookup = json.load(f)
|
||||
|
||||
# ── Config ─────────────────────────────────────────────────────────────────
|
||||
TOP_K = 6
|
||||
|
||||
SYSTEM_PROMPT = """You are an elite AP US History tutor. Your only goal is to help the student master APUSH and score a 5.
|
||||
|
||||
━━━ TOOL USE ━━━
|
||||
ALWAYS call search_textbook before answering any history question — no exceptions.
|
||||
For complex questions (LEQ/DBQ/thematic), call it 2-3 times with different search angles to get full coverage.
|
||||
|
||||
━━━ CITATIONS ━━━
|
||||
- Cite inline after every specific claim: (Ch5, p.153)
|
||||
- The **bolded sentences** in each source are the most relevant — prioritize citing and building on those
|
||||
- Never invent or guess a citation — if unsure, say "Outside textbook:"
|
||||
- If the textbook is silent on something relevant, supplement with general knowledge, clearly labeled
|
||||
|
||||
━━━ ACCURACY ━━━
|
||||
- Correct false premises immediately and directly — never reinforce a wrong assumption
|
||||
- Distinguish causation from correlation, primary from secondary causes
|
||||
- Note historiographical debates where relevant (e.g. revisionist vs traditional interpretations)
|
||||
- Be precise with dates, names, legislation, and turning points — vagueness loses points on the exam
|
||||
|
||||
━━━ FORMAT — match the question type exactly ━━━
|
||||
- Identification / one fact → one concise answer, one citation
|
||||
- SAQ (Short Answer) → 3 tight paragraphs: claim → evidence → analysis. No intro/conclusion fluff
|
||||
- LEQ (Long Essay) → Full essay: contextualization → thesis → 3 body paragraphs (each with specific evidence + analysis) → conclusion with complexity
|
||||
- DBQ → Same as LEQ plus: sourcing, audience/purpose/context for docs, corroboration across docs
|
||||
- Compare/contrast → Use parallel structure, explicit similarities AND differences
|
||||
- General question → Clear prose, as long as needed, no padding
|
||||
|
||||
━━━ APUSH EXAM SKILLS ━━━
|
||||
When writing essays, explicitly hit the College Board rubric:
|
||||
- Contextualization: zoom out to broader historical context BEFORE the thesis
|
||||
- Thesis: historically defensible, specific, addresses complexity (not just "there were many causes")
|
||||
- Evidence: at least 2 specific pieces of evidence per body paragraph
|
||||
- Analysis: explain HOW and WHY, not just what happened
|
||||
- Complexity: demonstrate nuance — turning points, continuity vs change, multiple causation, or cross-period connections
|
||||
|
||||
━━━ END EVERY RESPONSE WITH ━━━
|
||||
---
|
||||
**Sources Used:**
|
||||
[list each source: Ch# › Section › p.### — score: X.XXX]
|
||||
**Retrieval Confidence:** HIGH / MEDIUM / LOW
|
||||
**Exam Tip:** [one sentence of targeted advice for how this topic typically appears on the APUSH exam]"""
|
||||
TOP_K = 10
|
||||
|
||||
# ── Embed ──────────────────────────────────────────────────────────────────
|
||||
def embed_query(query: str) -> np.ndarray:
|
||||
@@ -89,11 +48,6 @@ def embed_query(query: str) -> np.ndarray:
|
||||
|
||||
# ── Highlight ──────────────────────────────────────────────────────────────
|
||||
def highlight_passage(query_emb: np.ndarray, passage: str) -> str:
|
||||
"""
|
||||
Bold the top 3 most query-relevant sentences using the already-loaded
|
||||
embedder. Reuses the query embedding computed during retrieval — zero
|
||||
extra model calls.
|
||||
"""
|
||||
sentences = [s.strip() for s in re.split(r'(?<=[.!?])\s+', passage) if len(s.strip()) > 20]
|
||||
if not sentences:
|
||||
return passage
|
||||
@@ -105,14 +59,13 @@ def highlight_passage(query_emb: np.ndarray, passage: str) -> str:
|
||||
show_progress_bar=False,
|
||||
)
|
||||
|
||||
scores = sent_embs @ query_emb # cosine sim (both normalized)
|
||||
scores = sent_embs @ query_emb
|
||||
top_n = min(3, len(scores))
|
||||
threshold = float(sorted(scores)[-top_n])
|
||||
|
||||
highlighted = passage
|
||||
for sent, score in zip(sentences, scores):
|
||||
if float(score) >= threshold:
|
||||
# avoid double-bolding if somehow already bolded
|
||||
if f"**{sent}**" not in highlighted:
|
||||
highlighted = highlighted.replace(sent, f"**{sent}**")
|
||||
|
||||
@@ -120,7 +73,7 @@ def highlight_passage(query_emb: np.ndarray, passage: str) -> str:
|
||||
|
||||
# ── Retrieve ───────────────────────────────────────────────────────────────
|
||||
def retrieve(query: str) -> dict:
|
||||
query_emb = embed_query(query) # compute once, reuse for highlighting
|
||||
query_emb = embed_query(query)
|
||||
|
||||
hits = qdrant.query_points(
|
||||
collection_name=COLLECTION,
|
||||
@@ -134,12 +87,7 @@ def retrieve(query: str) -> dict:
|
||||
).points
|
||||
|
||||
top_score = hits[0].score if hits else 0
|
||||
if top_score >= 0.70:
|
||||
confidence = "HIGH"
|
||||
elif top_score >= 0.50:
|
||||
confidence = "MEDIUM"
|
||||
else:
|
||||
confidence = "LOW"
|
||||
confidence = "HIGH" if top_score >= 0.70 else "MEDIUM" if top_score >= 0.50 else "LOW"
|
||||
|
||||
seen_parents = set()
|
||||
unique_hits = []
|
||||
@@ -149,14 +97,14 @@ def retrieve(query: str) -> dict:
|
||||
seen_parents.add(pid)
|
||||
unique_hits.append(h)
|
||||
|
||||
unique_hits = unique_hits[:4]
|
||||
unique_hits = unique_hits[:5]
|
||||
|
||||
sources = []
|
||||
for h in unique_hits:
|
||||
pid = h.payload["parent_id"]
|
||||
parts = parent_lookup.get(pid, [])
|
||||
full_text = "\n\n".join(p["text"] for p in parts)
|
||||
highlighted = highlight_passage(query_emb, full_text) # reuse query_emb
|
||||
pid = h.payload["parent_id"]
|
||||
parts = parent_lookup.get(pid, [])
|
||||
full_text = "\n\n".join(p["text"] for p in parts)
|
||||
highlighted = highlight_passage(query_emb, full_text)
|
||||
|
||||
sources.append({
|
||||
"score": h.score,
|
||||
@@ -174,7 +122,7 @@ def retrieve(query: str) -> dict:
|
||||
"sources": sources,
|
||||
}
|
||||
|
||||
# ── Origin bypass middleware ────────────────────────────────────────────────
|
||||
# ── Origin bypass middleware ───────────────────────────────────────────────
|
||||
class AllowAllOriginsMiddleware(BaseHTTPMiddleware):
|
||||
async def dispatch(self, request: Request, call_next):
|
||||
request._headers = request.headers.mutablecopy()
|
||||
@@ -182,15 +130,16 @@ class AllowAllOriginsMiddleware(BaseHTTPMiddleware):
|
||||
return await call_next(request)
|
||||
|
||||
# ── MCP Server ─────────────────────────────────────────────────────────────
|
||||
mcp = FastMCP("APUSH Tutor", instructions=SYSTEM_PROMPT)
|
||||
mcp = FastMCP("APUSH Tutor")
|
||||
|
||||
@mcp.tool()
|
||||
def search_textbook(query: str) -> str:
|
||||
"""
|
||||
Search the AP US History textbook for relevant passages.
|
||||
Use this for any question about US history before answering.
|
||||
Always cite sources inline and list all sources at the end.
|
||||
Bold or emphasize the most important phrases in your answer.
|
||||
Call this before answering ANY US history question.
|
||||
For broad topics call it multiple times with different search angles.
|
||||
Returns passages with the most relevant sentences bolded.
|
||||
Always cite inline (Ch#, p.###) and list sources at the end.
|
||||
"""
|
||||
retrieved = retrieve(query)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user