Spaces:
Sleeping
Sleeping
File size: 9,728 Bytes
dda5804 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 |
"""
Pinecone-backed vectorstore utilities for the AI Litigation Tracker.
Responsibilities:
- Embed case text using OpenAI embeddings.
- Chunk long documents and mean-pool embeddings into a single case vector.
- Upsert vectors into a Pinecone index.
- Run global similarity search for RAG (query_global).
- Look up a single case by normalized docket number or case name (get_case_by_filter).
Metadata stored with each vector includes:
- docket_number, case_name
- court_id, filing_date
- jurisdiction, courtlistener_url, latest_update
- n_docs (document count for the case)
"""
import os
import hashlib
import time
from typing import Dict, List, Optional
from dotenv import load_dotenv
from pinecone import Pinecone
from openai import OpenAI
import tiktoken
# Load environment variables (e.g., OPENAI_API_KEY, PINECONE_API_KEY, PINECONE_INDEX)
load_dotenv()
OPENAI_EMBED_MODEL = os.getenv("OPENAI_EMBED_MODEL", "text-embedding-3-small")
PINECONE_INDEX = os.getenv("PINECONE_INDEX", "ai-litigation-cases")
# Reuse single clients across calls
_oai = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
_pc = Pinecone(api_key=os.environ["PINECONE_API_KEY"])
try:
_index = _pc.Index(PINECONE_INDEX)
except Exception as e:
# Fail early with a clear message if the index has not been created yet.
raise RuntimeError(
f"Pinecone index '{PINECONE_INDEX}' not found. "
"Run data_updating_scripts.create_pinecone_index first."
) from e
# ============================================================
# Internal helpers
# ============================================================
def _norm(s: Optional[str]) -> Optional[str]:
"""
Normalize strings for case-insensitive lookups.
Returns lowercased, stripped text or None if input is not a string.
"""
return s.lower().strip() if isinstance(s, str) else None
def _make_id(court_id: str, docket_number: str) -> str:
"""
Create a stable, opaque ID for a (court_id, docket_number) pair.
Uses SHA-1 for compactness and to avoid leaking the raw identifiers in the ID.
"""
return hashlib.sha1(f"{court_id}|{docket_number}".encode()).hexdigest()
def _chunk_text(txt: str, max_tokens: int = 750) -> List[str]:
"""
Split text into chunks based on token count using tiktoken.
Args:
txt: Raw text to tokenize and chunk.
max_tokens: Maximum tokens per chunk (approximate prompt size control).
Returns:
A list of decoded text chunks. Returns [""] for empty input.
"""
enc = tiktoken.get_encoding("cl100k_base")
ids = enc.encode(txt or "")
chunks = [enc.decode(ids[i : i + max_tokens]) for i in range(0, len(ids), max_tokens)]
return chunks or [""]
def _embed_batch(texts: List[str]) -> List[List[float]]:
"""
Embed a batch of texts using the configured OpenAI embedding model.
Implements simple retry with backoff for transient errors.
Args:
texts: List of strings to embed.
Returns:
List of embedding vectors (one per input string), in order.
"""
out: List[List[float]] = []
i = 0
while i < len(texts):
batch = texts[i : i + 32] # modest batch size for reliability
for attempt in range(4):
try:
resp = _oai.embeddings.create(model=OPENAI_EMBED_MODEL, input=batch)
out.extend([d.embedding for d in resp.data])
break
except Exception:
if attempt == 3:
# After 4 attempts, re-raise to surface the failure.
raise
# Simple linear backoff to avoid hammering the API.
time.sleep(1.5 * (attempt + 1))
i += 32
return out
def _mean_pool(vectors: List[List[float]]) -> List[float]:
"""
Compute the element-wise mean of a list of vectors.
Used to aggregate chunk-level embeddings into a single case-level embedding.
"""
if not vectors:
return []
d = len(vectors[0])
acc = [0.0] * d
for v in vectors:
for j in range(d):
acc[j] += v[j]
return [x / len(vectors) for x in acc]
# ============================================================
# Public API: vector creation + indexing
# ============================================================
def case_to_vector_payload(
*,
docket_number: str,
case_name: str,
court_id: str,
filing_date: Optional[str],
concatenated_plain_text: str,
extra_meta: Optional[Dict] = None,
):
"""
Build a Pinecone-ready vector payload for a single case.
Steps:
1. Tokenize and chunk the concatenated case text.
2. Embed each chunk with OpenAI embeddings.
3. Mean-pool chunk embeddings into a single centroid vector.
4. Construct a stable ID and attach metadata used for filtering and display.
Args:
docket_number: Raw docket number string.
case_name: Case name/title.
court_id: CourtListener slug (e.g., "mdd", "nysd").
filing_date: Filing date as a string (e.g., "08302023" or "2023-08-30").
concatenated_plain_text: Combined plain text for all documents in the case.
extra_meta: Optional dict for additional metadata such as:
- n_docs
- courtlistener_url
- jurisdiction
- latest_update
Returns:
(stable_id, embedding_vector, metadata_dict)
"""
chunks = _chunk_text(concatenated_plain_text)
embs = _embed_batch(chunks)
centroid = _mean_pool(embs)
stable_id = _make_id(court_id, docket_number)
extra_meta = extra_meta or {}
metadata = {
"docket_number": docket_number,
"docket_number_norm": _norm(docket_number),
"case_name": case_name,
"case_name_norm": _norm(case_name),
"court_id": court_id,
"filing_date": filing_date or "",
"n_docs": extra_meta.get("n_docs"),
"courtlistener_url": extra_meta.get("courtlistener_url"),
"jurisdiction": extra_meta.get("jurisdiction"),
"latest_update": extra_meta.get("latest_update"),
}
# Drop None values so the index metadata stays lean.
metadata = {k: v for k, v in metadata.items() if v is not None}
return stable_id, centroid, metadata
def already_indexed(*, court_id: str, docket_number: str) -> bool:
"""
Check whether a case is already present in the Pinecone index.
Args:
court_id: CourtListener slug.
docket_number: Docket number for the case.
Returns:
True if an entry with the stable ID exists in the index, else False.
"""
vid = _make_id(court_id, docket_number)
res = _index.fetch(ids=[vid])
vectors = (res or {}).get("vectors") or {}
return vid in vectors
def upsert_cases(vectors: List[Dict]) -> None:
"""
Upsert a list of vector payloads into the Pinecone index.
Args:
vectors: List of dicts of the form {"id": ..., "values": [...], "metadata": {...}}.
Notes:
- Pinecone handles deduplication by ID, so upserts are idempotent.
- We still batch requests for efficiency and API friendliness.
"""
for i in range(0, len(vectors), 100):
_index.upsert(vectors=vectors[i : i + 100])
# ============================================================
# Public API: querying
# ============================================================
def query_global(question: str, top_k: int = 5) -> List[Dict]:
"""
Run a global semantic search over all cases for a natural language question.
Args:
question: Text query to embed and search with.
top_k: Maximum number of matches to return.
Returns:
A list of dictionaries for each match:
{
"score": <similarity score>,
...<all stored metadata fields>...
}
"""
q = _embed_batch([question])[0]
res = _index.query(vector=q, top_k=top_k, include_metadata=True)
return [{"score": m["score"], **m["metadata"]} for m in res.get("matches", [])]
def get_case_by_filter(
*,
docket_number: Optional[str] = None,
case_name: Optional[str] = None,
) -> Optional[Dict]:
"""
Look up a single case by normalized docket number and/or case name.
This is primarily used by case-specific Q&A in rag/chains.py.
Args:
docket_number: Optional docket string (case-insensitive).
case_name: Optional case name string (case-insensitive).
Returns:
A metadata dict for the first matching case, or None if no match.
The returned metadata mirrors what was stored in case_to_vector_payload.
"""
flt: Dict[str, str] = {}
if docket_number:
flt["docket_number_norm"] = _norm(docket_number)
if case_name:
flt["case_name_norm"] = _norm(case_name)
if not flt:
return None
# We do a "dummy" query using a zero vector and rely on metadata filter
# to select the matching case. The dimension must match the embed model.
dim = 1536 if "3-small" in OPENAI_EMBED_MODEL else 3072
res = _index.query(
vector=[0.0] * dim,
top_k=1,
include_metadata=True,
filter=flt,
)
matches = res.get("matches", [])
return matches[0]["metadata"] if matches else None |